{
  "listings": [
    {
      "slug": "glm-4-7-flash",
      "name": "GLM-4.7-Flash",
      "company": "JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai)",
      "grade": "B",
      "fit": "yes",
      "why": "Card names glm47 and glm45 parsers, both in vLLM 0.33.0; Ollama's glm-4.7-flash (19 GB) carries tools and thinking badges; acc-20 found it kept progressing where Qwen 3.6 looped. MLA keeps 64K of cache near 3.5 GB.",
      "source": "https://huggingface.co/zai-org/GLM-4.7-Flash",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/glm-4-7-flash"
    },
    {
      "slug": "qwen3-8-27b",
      "name": "Qwen3.8-27B",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "B",
      "fit": "yes",
      "why": "acc-39 ran OpenCode on UD-Q6_K_XL and passed 77 of 77 hidden tests with 97 percent prompt-cache hits; the SGLang cookbook names qwen3_coder and qwen3 parsers and enable_thinking false.",
      "source": "https://huggingface.co/Qwen/Qwen3.8-27B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-8-27b"
    },
    {
      "slug": "agents-a1-4b",
      "name": "Agents-A1-4B",
      "company": "Shanghai Artificial Intelligence Laboratory (the InternAgent team there, per the card's recommended system prompt and the template's default prompt); the GitHub LICENSE leaves the copyright line as the Apache template",
      "grade": "C",
      "fit": "caveat",
      "why": "Card names qwen3_coder and qwen3 parsers; Q4_K_M GGUF is 2.7 GB. Qwen3.5 hybrid attention re-processes context each turn (acc-03) and the qwen3_coder parser drops streamed calls to text (acc-29); no account runs this 4B in a harness.",
      "source": "https://huggingface.co/InternScience/Agents-A1-4B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/agents-a1-4b"
    },
    {
      "slug": "apertus-v1-5-8b",
      "name": "Apertus v1.5 8B",
      "company": "ETH Zurich and EPFL (the Swiss National AI Institute, named in the Acceptable Use Policy)",
      "grade": "C",
      "fit": "caveat",
      "why": "Architecture Apertus1p5ForConditionalGeneration is absent from vLLM 0.33.0's registry and llama.cpp's apertus entry is the 1.0 model, so the only serving path is the publisher's Docker fork; the gated card could not be read.",
      "source": "https://huggingface.co/api/models/swiss-ai/Apertus-v1.5-8B?blobs=true",
      "confidence": "low",
      "url": "https://www.anchorterminal.com/tools/apertus-v1-5-8b"
    },
    {
      "slug": "apodex-1-1-mini",
      "name": "Apodex-1.1-mini",
      "company": "Apodex (the site's copyright line reads 2026 Apodex)",
      "grade": "C",
      "fit": "caveat",
      "why": "Card names qwen3_coder and qwen3 parsers and a 21.7 GB Q4_K_M GGUF with llama-server --jinja fits 32 GB. The Qwen3.5 MoE base re-processes context on hybrids (acc-03), loops in agents (acc-01, acc-15) and qwen3_coder drops streamed calls (acc-29).",
      "source": "https://huggingface.co/apodex/Apodex-1.1-mini",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/apodex-1-1-mini"
    },
    {
      "slug": "bigbang-v1",
      "name": "BigBang-v1",
      "company": "The Endless Frontier (endless-frontier)",
      "grade": "C",
      "fit": "caveat",
      "why": "Card gives vLLM and SGLang commands with qwen3_coder and qwen3 parsers and asks for 128K context so thinking is not cut short; the publisher ships BF16 only (71.9 GB) and a third-party Q4_K_M (21.9 GB) is the local route. Qwen3.6 hybrid re-processing (acc-03) applies.",
      "source": "https://huggingface.co/endless-frontier/BigBang-v1",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/bigbang-v1"
    },
    {
      "slug": "darwin-36b-opus",
      "name": "Darwin-36B-Opus",
      "company": "VIDRAFT Inc. (named on vidraft.net); the card's citation names FINAL-Bench and VIDRAFT_LAB",
      "grade": "C",
      "fit": "caveat",
      "why": "Publisher's POCKET Q4_K_M (21.2 GB) runs on stock llama.cpp (qwen35moe) and the template carries the Qwen3.6 tool format, but the card names no parser, says responses always open with a think trace, and tool calling is undocumented.",
      "source": "https://huggingface.co/FINAL-Bench/POCKET-35B-GGUF",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/darwin-36b-opus"
    },
    {
      "slug": "deepseek-v4-flash-0731",
      "name": "DeepSeek-V4-Flash-0731",
      "company": "DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)",
      "grade": "C",
      "fit": "caveat",
      "why": "acc-08 drives pi.dev with it at IQ2 on a 128 GB M5 Max at 30 to 35 tokens per second; acc-40 got 11 tokens per second and Chinese replies at IQ3_XXS on Strix Halo. Unsloth's 3-bit is 103 GB; no Jinja template ships.",
      "source": "https://news.ycombinator.com/item?id=48555993",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/deepseek-v4-flash-0731"
    },
    {
      "slug": "gemma-4-12b-it",
      "name": "Gemma 4 12B IT",
      "company": "Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author)",
      "grade": "BB",
      "fit": "caveat",
      "why": "Q4_0 QAT is 7 GB with a gemma4 tool parser in vLLM and tools badge on Ollama; acc-24 got 100 percent tool calls on an 8 GB GPU. acc-09 reports Gemma stopping mid tool call and acc-27 an Ollama parser bug on quoted arguments, fixed in 0.20.2.",
      "source": "https://huggingface.co/api/models/google/gemma-4-12B-it-qat-q4_0-gguf?blobs=true",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/gemma-4-12b-it"
    },
    {
      "slug": "gemma-4-26b-a4b-it",
      "name": "Gemma 4 26B A4B IT",
      "company": "Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author)",
      "grade": "BB",
      "fit": "caveat",
      "why": "Q4_0 QAT (14.4 GB) and gemma4 parsers are in place, but acc-33 saw 3 of 3 automated runs fail, acc-34 20 of 33 with early give-up, acc-38 incomplete code and acc-35 112 second turns; acc-42 found vLLM lacked a 4-bit MoE path.",
      "source": "https://huggingface.co/api/models/google/gemma-4-26B-A4B-it-qat-q4_0-gguf?blobs=true",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it"
    },
    {
      "slug": "gemma-4-31b-it",
      "name": "Gemma 4 31B IT",
      "company": "Google LLC (the registrant of gemma.ai; the card names Google DeepMind as author)",
      "grade": "BB",
      "fit": "caveat",
      "why": "acc-37 found gemma4 31b NVFP4 in OpenCode about 50 percent faster than Qwen with slightly better output; Q4_0 QAT is 17.6 GB and KV at 64K is large (16 of 60 layers global). acc-09 saw Gemma stop mid tool call and acc-27 an Ollama parser bug since fixed.",
      "source": "https://pac.commonsware.com/archive/mlx-and-nvfp4-ftw/",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/gemma-4-31b-it"
    },
    {
      "slug": "gemma-4-e2b-it",
      "name": "Gemma 4 E2B IT",
      "company": "Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author)",
      "grade": "C",
      "fit": "caveat",
      "why": "Tool calling and the gemma4 parser work at 3.35 GB Q4_0, but acc-45 found E2B broke React imports and Google's own card puts it far below the larger sizes on coding; usable only for trivial edits.",
      "source": "https://huggingface.co/api/models/google/gemma-4-E2B-it-qat-q4_0-gguf?blobs=true",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/gemma-4-e2b-it"
    },
    {
      "slug": "gemma-4-e4b-it",
      "name": "Gemma 4 E4B IT",
      "company": "Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author)",
      "grade": "B",
      "fit": "caveat",
      "why": "acc-45 saw E4B pass 4 of 4 Python tasks and a six-step agent task; Q4_0 is 5.15 GB with a gemma4 parser. The card's agent and coding figures sit far below the 12B and up, so scoped edits only.",
      "source": "https://www.kartit.net/blog/gemma4-local-benchmark",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/gemma-4-e4b-it"
    },
    {
      "slug": "glm-4-5-air",
      "name": "GLM-4.5-Air",
      "company": "JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai, per the first batch's check on 21 October 2026)",
      "grade": "C",
      "fit": "caveat",
      "why": "glm45 tool and reasoning parsers are in vLLM 0.33.0 and the template turns thinking off; Z.ai ships BF16 and FP8 only (112 GB), so a 128 GB box runs Unsloth's UD-Q4_K_XL (68 GB) under llama.cpp's glm4moe. No account, 128K context cap.",
      "source": "https://huggingface.co/zai-org/GLM-4.5-Air",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/glm-4-5-air"
    },
    {
      "slug": "glm-5-3-flash",
      "name": "GLM-5.3-Flash",
      "company": "Z.AI Co., Ltd (named in the licence's copyright line)",
      "grade": "B",
      "fit": "caveat",
      "why": "Unsloth's 1-bit (93 GB) and 3-bit (120 GB) GGUFs fit 128 GB and glm47 parsers are in vLLM 0.33.0; acc-22 ran it at Q2 on 192 GB as a reviewer and found both configurations too slow for an interactive coding session.",
      "source": "https://unsloth.ai/docs/models/glm-5.3-flash",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/glm-5-3-flash"
    },
    {
      "slug": "gpt-oss-120b",
      "name": "gpt-oss-120b",
      "company": "OpenAI (named as publisher on the card; the openai-api listing's terms name OpenAI OpCo, LLC)",
      "grade": "BB",
      "fit": "caveat",
      "why": "63 GB MXFP4 fits an 80 GB GPU or 128 GB box (acc-21 and acc-40 loaded it on Strix Halo) and vLLM has the openai parser; but one tool call per turn, acc-30 shows parallel calls broken by a repository commit, and reasoning cannot be turned off.",
      "source": "https://huggingface.co/openai/gpt-oss-120b",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/gpt-oss-120b"
    },
    {
      "slug": "gpt-oss-20b",
      "name": "gpt-oss-20b",
      "company": "OpenAI (named in USAGE_POLICY and the model card; the LICENSE file's copyright line is the unfilled Apache template)",
      "grade": "B",
      "fit": "caveat",
      "why": "12 GB MXFP4 fits 16 GB and Ollama's gpt-oss:20b carries tools and thinking, but acc-47 hit AI_NoSuchToolError in OpenCode and loops, one tool call per turn is the design, and reasoning only drops to low.",
      "source": "https://huggingface.co/openai/gpt-oss-20b",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/gpt-oss-20b"
    },
    {
      "slug": "granite-4-2-3b",
      "name": "Granite 4.2 3B",
      "company": "International Business Machines Corporation (per IBM's naming guidance, as read for granite-guardian)",
      "grade": "C",
      "fit": "caveat",
      "why": "The card documents OpenCode and Pi configurations, names qwen3_coder and granite parsers (built into vLLM 0.33.0) and ships a 2.2 GB Q4_K_M; a 3.7B model will not finish most edit-and-test tasks, and the qwen3_coder parser drops streamed calls (acc-29).",
      "source": "https://huggingface.co/ibm-granite/granite-4.2-3b",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/granite-4-2-3b"
    },
    {
      "slug": "hy3",
      "name": "Hy3",
      "company": "Tencent (named in the LICENSE file's copyright line)",
      "grade": "B",
      "fit": "caveat",
      "why": "hy_v3 tool and reasoning parsers are in vLLM 0.33.0 and llama.cpp has hy_v3; only AngelSlim's IQ1_M (89 GB) fits a 128 GB box, and no account tests tool calls at 1 bit.",
      "source": "https://huggingface.co/tencent/Hy3",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/hy3"
    },
    {
      "slug": "inkling-small",
      "name": "Inkling-Small",
      "company": "Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary)",
      "grade": "B",
      "fit": "caveat",
      "why": "Unsloth's UD-Q2_K_XL (88 GB) fits 128 GB and inkling parsers are in vLLM 0.33.0, but the publisher's NVFP4 needs 170 GB on Blackwell, the GGUF architecture is not in llama.cpp master's list, and no account runs it.",
      "source": "https://huggingface.co/api/models/unsloth/Inkling-Small-GGUF?blobs=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/inkling-small"
    },
    {
      "slug": "kolibri-1",
      "name": "Kolibri-1",
      "company": "Aleph Alpha GmbH",
      "grade": "C",
      "fit": "caveat",
      "why": "vLLM 0.33.0 loads it natively with hf reasoning and tool parsers, llama.cpp has kolibri1 and bartowski's Q4_K_M is 49.8 GB for a 64 GB machine; the hf parser lacks tool_choice required and strict tools, Aleph Alpha ships no GGUF, and no account runs it in a harness.",
      "source": "https://huggingface.co/Aleph-Alpha/Kolibri-1",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/kolibri-1"
    },
    {
      "slug": "laguna-s-2-1",
      "name": "Laguna S 2.1",
      "company": "Poolside, Inc. (named in Poolside's Terms of Use)",
      "grade": "C",
      "fit": "caveat",
      "why": "Fits only the 128 GB tier: Q4_K_M is 96 GB plus about 12 GiB of KV at 64K. Accounts report 20 to 28 tokens per second on Strix Halo; discussions report thinking loops and Cline failing to parse its tool calls.",
      "source": "https://huggingface.co/api/models/poolside/Laguna-S-2.1-GGUF/tree/main?recursive=true",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/laguna-s-2-1"
    },
    {
      "slug": "laguna-xs-2-1",
      "name": "Laguna XS 2.1",
      "company": "Poolside, Inc. (named in Poolside's Terms of Use)",
      "grade": "C",
      "fit": "caveat",
      "why": "Q4_K_M 20.3 GB plus about 10 GiB KV at 64K fits 32 GB and up, with poolside_v1 parsers and an Ollama tools template. The card warns Ollama on Apple Silicon may return empty output and vLLM drops newline-free calls without PR 47311. One user runs it with Pi.",
      "source": "https://huggingface.co/poolside/Laguna-XS-2.1/raw/main/README.md",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/laguna-xs-2-1"
    },
    {
      "slug": "ling-3-0-tiny",
      "name": "Ling-3.0-tiny",
      "company": "Ant Intelligence (Hang Zhou) Technology Co., Ltd (named as provider in the training-content summary); inclusionAI, Ant Group in the site footer of inclusion-ai.org",
      "grade": "C",
      "fit": "caveat",
      "why": "4.8 GB Q4_K_M fits anything and vLLM names ling3 parsers, but 1.3B active parameters are thin for a harness, the KDA hybrid may re-process context, the vLLM route needs trust_remote_code and a discussion reports looping.",
      "source": "https://huggingface.co/inclusionAI/Ling-3.0-tiny/raw/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/ling-3-0-tiny"
    },
    {
      "slug": "mimo-v2-6-distill-qwen-9b",
      "name": "MiMo-V2.6 Distill Qwen 9B",
      "company": "Xiaomi (the card and report name Xiaomi MiMo; no LICENSE file names a copyright holder)",
      "grade": "C",
      "fit": "caveat",
      "why": "Ollama's 6.6 GB q4_K_M fits 16 GB with tools and thinking off, but Xiaomi ships BF16 only, names no tool parser on the card, and discussions report unreliable tool calls and runaway generation on llama.cpp.",
      "source": "https://ollama.com/library/mimo-v2.6/tags",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/mimo-v2-6-distill-qwen-9b"
    },
    {
      "slug": "minimax-m2-5",
      "name": "MiniMax-M2.5",
      "company": "MiniMax AI (named in the licences; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)",
      "grade": "C",
      "fit": "caveat",
      "why": "Only the 128 GB tier fits, at 3 bits or below (Unsloth UD-Q3_K_XL 101 GB plus about 15 GiB KV at 64K); the template forces a think block on every turn and vLLM needs trust_remote_code.",
      "source": "https://huggingface.co/api/models/unsloth/MiniMax-M2.5-GGUF/tree/main?recursive=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/minimax-m2-5"
    },
    {
      "slug": "minimax-m2-7",
      "name": "MiniMax-M2.7",
      "company": "MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)",
      "grade": "C",
      "fit": "caveat",
      "why": "minimax_m2 tool parser is named, but only a 128 GB box fits, at 3 bits (UD-IQ3_XXS 80 GB plus about 16 GB KV at 64K); thinking opens every turn, vLLM needs trust_remote_code, and commercial use needs written authorisation.",
      "source": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/main/docs/vllm_deploy_guide.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/minimax-m2-7"
    },
    {
      "slug": "muse-glimmer-30b",
      "name": "Muse Glimmer 30B",
      "company": "Meta (the verified Hugging Face organisation is \"Meta Inc.\", the card names Meta Superintelligence Lab, and the licence carries no copyright line)",
      "grade": "BB",
      "fit": "caveat",
      "why": "Named parsers (vLLM muse_glimmer), Ollama tools badge, 17 GB Q4_K_M with 128K, card measures 75 tok/s on an RTX 5090 and 24 on an M4 Max; but one tool call per turn, thinking always on, and no account.",
      "source": "https://dev.meta.ai/docs/muse-glimmer/deploy.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/muse-glimmer-30b"
    },
    {
      "slug": "nanbeige-4-3-33b",
      "name": "Nanbeige4.3-33B",
      "company": "BOSS Zhipin (Kanzhun Limited), unconfirmed",
      "grade": "C",
      "fit": "caveat",
      "why": "Card names qwen3_coder and qwen3 parsers and preserve_thinking for tool loops; bartowski's Q4_K_M is 20.8 GB plus about 14 GB KV at 64K, so 32 GB-plus. Dense 33B decode at 32K is unmeasured and no account runs a harness.",
      "source": "https://huggingface.co/Nanbeige/Nanbeige4.3-33B/raw/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/nanbeige-4-3-33b"
    },
    {
      "slug": "nemotron-3-5-lightning-30b-a3b",
      "name": "Nemotron 3.5 Lightning 30B A3B",
      "company": "NVIDIA Corporation",
      "grade": "B",
      "fit": "caveat",
      "why": "qwen3_coder and nemotron_v3 parsers on a tagged vLLM, force_nonempty_content for coding agents, 18.9 GB Q4_0 with only 6 attention layers, so 64K adds under 1 GB; but it is a Mamba hybrid, where accounts report full re-processing each turn, and no account runs it.",
      "source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16/raw/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/nemotron-3-5-lightning-30b-a3b"
    },
    {
      "slug": "nemotron-3-nano-30b-a3b",
      "name": "Nemotron 3 Nano 30B A3B",
      "company": "NVIDIA Corporation",
      "grade": "B",
      "fit": "caveat",
      "why": "qwen3_coder parser named and SWE-Bench (OpenHands) 38.8 on the card; Ollama's 18 GB to 24 GB build with tools fits 32 GB. But vLLM needs trust_remote_code plus a parser file from the repository, and it is a Mamba hybrid with no account timing turn two.",
      "source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/raw/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/nemotron-3-nano-30b-a3b"
    },
    {
      "slug": "nemotron-3-super-120b-a12b",
      "name": "Nemotron 3 Super 120B A12B",
      "company": "NVIDIA Corporation",
      "grade": "B",
      "fit": "caveat",
      "why": "qwen3_coder parser and force_nonempty_content for agents; 4-bit fits a 128 GB box (Ollama 87 GB). SEP's Strix Halo account got 17 tok/s in VS Code Copilot, under the 20 target, with ROCm out of memory and Vulkan lock-ups.",
      "source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/raw/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/nemotron-3-super-120b-a12b"
    },
    {
      "slug": "occamy-1-0",
      "name": "occamy-1.0",
      "company": "Accio (Accio-Lab)",
      "grade": "C",
      "fit": "caveat",
      "why": "qwen3_coder and qwen3 parsers, preserve_thinking for tool loops, trained in OpenClaw and Hermes; Q4_K_M 21.2 GB fits 24 to 32 GB. But it keeps the Qwen3.6 hybrid layers that accounts say re-process every turn, GGUFs need NFC input, and no account runs Occamy.",
      "source": "https://huggingface.co/Accio-Lab/occamy-1.0/raw/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/occamy-1-0"
    },
    {
      "slug": "ornith-1-5-35b-a3b",
      "name": "Ornith-1.5-35B-A3B",
      "company": "Ornith",
      "grade": "C",
      "fit": "caveat",
      "why": "Q4_K_M 21.7 GB plus 1.2 GiB KV fits 24 GB and up with qwen3_xml or qwen3_coder parsers and an Ollama tools tag, but a 23-comment discussion says it fails basic file read and write tool calls, another reports constant looping, and the hybrid re-processes context.",
      "source": "https://huggingface.co/api/models/ornith-ai/Ornith-1.5-35B-A3B-GGUF/tree/main?recursive=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/ornith-1-5-35b-a3b"
    },
    {
      "slug": "ornith-1-5-9b",
      "name": "Ornith-1.5-9B",
      "company": "Ornith",
      "grade": "C",
      "fit": "caveat",
      "why": "Q4_K_M 5.8 GB plus 2 GiB KV fits any GPU with qwen3_xml or qwen3_coder parsers and an Ollama tools tag, but it reasons at length by default (a discussion says very long), the switch is template-only, and the hybrid re-processes context.",
      "source": "https://huggingface.co/api/models/ornith-ai/Ornith-1.5-9B-GGUF/tree/main?recursive=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/ornith-1-5-9b"
    },
    {
      "slug": "qwen3-30b-a3b-instruct-2507",
      "name": "Qwen3-30B-A3B-Instruct-2507",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "C",
      "fit": "caveat",
      "why": "Never thinks, 3B active, plain attention so the cache carries across turns; qwen3:30b-instruct is 19 GB plus about 6.4 GB KV at 64K, fitting 32 GB. The card names Qwen-Agent, not a vLLM parser, and no account runs a harness on this edition.",
      "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-30b-a3b-instruct-2507"
    },
    {
      "slug": "qwen3-4b-instruct-2507",
      "name": "Qwen3-4B-Instruct-2507",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "C",
      "fit": "caveat",
      "why": "Hermes-style JSON tool calls in the template and vLLM's hermes parser, but the card names no parser and no account runs a harness on it; a 4B general model, 2.5 GB in Ollama.",
      "source": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-4b-instruct-2507"
    },
    {
      "slug": "qwen3-5-122b-a10b",
      "name": "Qwen3.5-122B-A10B",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "B",
      "fit": "caveat",
      "why": "Parsers named and thinking off per request; Ollama's 81 GB Q4 build needs a 128 GB machine, acc-01 keeps it for harder tasks, acc-49 waited over five minutes for one answer.",
      "source": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-5-122b-a10b"
    },
    {
      "slug": "qwen3-5-4b",
      "name": "Qwen3.5-4B",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "C",
      "fit": "caveat",
      "why": "Card names qwen3_coder and qwen3 parsers and enable_thinking false; 2.7 GB at Q4 decodes fast on 16 GB, but no account runs a harness on the 4B and the family's smaller sizes loop.",
      "source": "https://huggingface.co/Qwen/Qwen3.5-4B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-5-4b"
    },
    {
      "slug": "qwen3-5-9b",
      "name": "Qwen3.5-9B",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "C",
      "fit": "caveat",
      "why": "acc-34 scored 27 of 33 coding tasks on a 16 GB GPU and acc-05 finished an OpenCode task offline, but acc-28 saw Ollama print the bash call as text, halting the agent.",
      "source": "https://huggingface.co/Qwen/Qwen3.5-9B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-5-9b"
    },
    {
      "slug": "qwen3-6-27b",
      "name": "Qwen3.6-27B",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "B",
      "fit": "caveat",
      "why": "acc-09 and acc-18 call it solid on a 3090 and two V100s, but acc-32 saw the qwen3_coder parser fail on \u003c \u003e and streamed payloads, acc-17 slow past 64K, acc-13 weak for code.",
      "source": "https://huggingface.co/Qwen/Qwen3.6-27B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-6-27b"
    },
    {
      "slug": "qwen3-6-35b-a3b",
      "name": "Qwen3.6-35B-A3B",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "B",
      "fit": "caveat",
      "why": "acc-34 passed 29 of 29 tasks and acc-19 saw 75 tok/s on a 16 GB GPU, but acc-01 and acc-15 report edit-tool loops and ignored directives, and acc-33 failed 5 of 7 runs on 48 GB.",
      "source": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-6-35b-a3b"
    },
    {
      "slug": "qwen3-8-flash-next",
      "name": "Qwen3.8-Flash-Next",
      "company": "Alibaba Cloud (Singapore) Private Limited, the registrant of qwen.ai. The licence names Qwen as copyright holder",
      "grade": "C",
      "fit": "caveat",
      "why": "Parsers named in the SGLang cookbook and thinking off per request, but Ollama's Linux Q4 build is 120 GB and acc-39's 93.7 GB UD-IQ4_XS would not fit beside a second model on 128 GB.",
      "source": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-8-flash-next"
    },
    {
      "slug": "qwen3-coder-30b-a3b-instruct",
      "name": "Qwen3-Coder-30B-A3B-Instruct",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "C",
      "fit": "caveat",
      "why": "Non-thinking coder with a shipped parser and vLLM's qwen3_xml entry, 19 GB in Ollama; acc-29 saw vLLM leave calls as text when streaming and acc-51 ran it as OpenCode's coder.",
      "source": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-coder-30b-a3b-instruct"
    },
    {
      "slug": "qwen3-coder-next",
      "name": "Qwen3-Coder-Next",
      "company": "Alibaba Cloud (the Qwen team's employer; no LICENSE file names a copyright holder for this model)",
      "grade": "C",
      "fit": "caveat",
      "why": "Built for coding agents, no thinking, publisher Q4_K_M GGUF at 48 GB and vLLM 0.15 parser; acc-33 got a correct implementation in 2.5 minutes before the runtime crashed and later output degraded.",
      "source": "https://huggingface.co/Qwen/Qwen3-Coder-Next",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-coder-next"
    },
    {
      "slug": "qwopus3-8-27b-flash-v2",
      "name": "Qwopus3.8-27B-Flash-V2",
      "company": "Jackrong, an individual Hugging Face account",
      "grade": "C",
      "fit": "caveat",
      "why": "Tuned for agent loops with GGUFs and an MTP head, but the template switches to Hermes-style JSON calls the card never documents, the card itself reports loops, and no account runs it.",
      "source": "https://huggingface.co/Jackrong/Qwopus3.8-27B-Flash-V2",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwopus3-8-27b-flash-v2"
    },
    {
      "slug": "recursive-flint",
      "name": "Recursive Flint",
      "company": "Recursive Superintelligence, Inc. (the Hugging Face organisation's full name)",
      "grade": "C",
      "fit": "caveat",
      "why": "Documented tool calls with flint parsers in the SGLang kit, llama.cpp --jinja, Ollama 23 GB and vLLM 0.33 support, but reasoning cannot be turned off and the card asks for 32K of output budget per hard task.",
      "source": "https://huggingface.co/RecursiveSI/flint",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/recursive-flint"
    },
    {
      "slug": "spark-x2-6-4b",
      "name": "Spark-X2.6-4B",
      "company": "XHToken (the licence's copyright line); the Hugging Face organisation's name is SparkLLM",
      "grade": "C",
      "fit": "caveat",
      "why": "Built for on-device agents with the spark25 parser in SGLang 0.5.22, llama.cpp b10828 --jinja and a 3.5 GB publisher GGUF, but a 4B model with no harness account and a template that errors on string arguments.",
      "source": "https://huggingface.co/XHToken/Spark-X2.6-4B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/spark-x2-6-4b"
    },
    {
      "slug": "swift-1-5-qwen3-8-27b",
      "name": "Swift 1.5 Qwen3.8-27B",
      "company": "UkisAI",
      "grade": "C",
      "fit": "caveat",
      "why": "Qwen3.8-27B's template and parsers (qwen3_coder, qwen3) with 17.4 GB Q4 GGUF and MTP, but UkisAI's licence stops commercial use at US$1M revenue and no account runs it.",
      "source": "https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/swift-1-5-qwen3-8-27b"
    },
    {
      "slug": "swift-1-5-qwen3-8-flash-next",
      "name": "Swift 1.5 Qwen3.8-Flash-Next",
      "company": "UkisAI",
      "grade": "C",
      "fit": "caveat",
      "why": "Only the 76 GB IQ3_XXS or 84.7 GB IQ3_S GSQ-RCO builds fit a 128 GB machine, below 4 bits; the card's two commands name different parsers and two licences apply.",
      "source": "https://huggingface.co/api/models/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF?blobs=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/swift-1-5-qwen3-8-flash-next"
    },
    {
      "slug": "thinkingcap-qwen3-8-27b",
      "name": "ThinkingCap (Qwen3.8 27B)",
      "company": "BottleCap AI (named in the licence's copyright line)",
      "grade": "C",
      "fit": "caveat",
      "why": "Qwen3.8-27B's parsers named on the card (qwen3_xml for vLLM, qwen3_coder for SGLang), ungated 17.4 GB Q4 GGUF with MTP, but PolyForm Small Business bars organisations of 100 people or US$1M, and no account.",
      "source": "https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.8-27B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/thinkingcap-qwen3-8-27b"
    },
    {
      "slug": "carbon-x",
      "name": "Carbon-X",
      "company": "Hugging Face (the HuggingFaceBio organisation, named Hugging Face Biology Research on the Hub)",
      "grade": "C",
      "fit": "no",
      "why": "A DNA model; the card lists no tool calling or coding training and recommends enable_thinking false. Not aimed at a harness.",
      "source": "https://huggingface.co/HuggingFaceBio/Carbon-X",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/carbon-x"
    },
    {
      "slug": "cosmos-reason2-2b",
      "name": "Cosmos-Reason2-2B",
      "company": "NVIDIA Corporation",
      "grade": "C",
      "fit": "no",
      "why": "A physical-AI video reasoner; NVIDIA documents no tool calling and the GitHub README's recipes are vLLM video prompts.",
      "source": "https://raw.githubusercontent.com/nvidia-cosmos/cosmos-reason2/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/cosmos-reason2-2b"
    },
    {
      "slug": "deepseek-v3-2",
      "name": "DeepSeek-V3.2",
      "company": "DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)",
      "grade": "C",
      "fit": "no",
      "why": "689.5 GB of FP8 and the smallest third-party GGUF is 161 GB (UD-TQ1_0); nothing fits a 16 to 128 GB machine.",
      "source": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V3.2?blobs=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/deepseek-v3-2"
    },
    {
      "slug": "deepseek-v4-1-flash",
      "name": "DeepSeek V4.1 Flash",
      "company": "DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)",
      "grade": "C",
      "fit": "no",
      "why": "The smallest GGUF is 233 GB (UD-IQ1_S) and Unsloth's table starts at 128 to 145 GB only with the Engram table left on SSD; a 16 to 128 GB machine cannot hold it.",
      "source": "https://huggingface.co/api/models/unsloth/DeepSeek-V4.1-Flash-1017-GGUF?blobs=true",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/deepseek-v4-1-flash"
    },
    {
      "slug": "diffusiongemma-26b-a4b-it",
      "name": "DiffusionGemma 26B A4B IT",
      "company": "Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author)",
      "grade": "D",
      "fit": "no",
      "why": "Tool calling is listed as a feature but no parser or serving command for it exists; Google's guide needs a GPU over 60 GB, and the Unsloth GGUF runs only on an unmerged llama.cpp pull request (24423).",
      "source": "https://huggingface.co/google/diffusiongemma-26B-A4B-it",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/diffusiongemma-26b-a4b-it"
    },
    {
      "slug": "glm-5-3",
      "name": "GLM-5.3",
      "company": "JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai). The licence names Z.AI",
      "grade": "B",
      "fit": "no",
      "why": "755.7 GB of FP8 and the smallest third-party GGUF is 217 GB (UD-IQ1_S); nothing fits 16 to 128 GB.",
      "source": "https://huggingface.co/api/models/zai-org/GLM-5.3?blobs=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/glm-5-3"
    },
    {
      "slug": "inkling",
      "name": "Inkling",
      "company": "Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary)",
      "grade": "B",
      "fit": "no",
      "why": "The smallest build is Unsloth's 1-bit at 270 GB; nothing fits 16 to 128 GB.",
      "source": "https://unsloth.ai/docs/models/inkling",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/inkling"
    },
    {
      "slug": "internvl3-5-gpt-oss-20b-a4b-preview",
      "name": "InternVL3.5-GPT-OSS-20B-A4B-Preview",
      "company": "Shanghai AI Laboratory (OpenGVLab), not named in the repository",
      "grade": "D",
      "fit": "no",
      "why": "The shipped chat template has no tools block and the card documents no tool calling; thinking is a system prompt, not a switch.",
      "source": "https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF/raw/main/chat_template.jinja",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/internvl3-5-gpt-oss-20b-a4b-preview"
    },
    {
      "slug": "kimi-k3",
      "name": "Kimi K3",
      "company": "Moonshot AI (the licence names Moonshot AI; the Kimi Platform terms name Moonshot AI PTE. LTD.)",
      "grade": "B",
      "fit": "no",
      "why": "1,561 GB of MXFP4 and 466 GB for the smallest third-party GGUF; nothing fits 16 to 128 GB.",
      "source": "https://huggingface.co/api/models/moonshotai/Kimi-K3?blobs=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/kimi-k3"
    },
    {
      "slug": "kronos",
      "name": "Kronos",
      "company": "No company. Yu Shi publishes it from the GitHub account shiyu-coder and the Hugging Face user account NeoQuasar (named ShiYu), with co-authors at Tsinghua University per the paper",
      "grade": "C",
      "fit": "no",
      "why": "Kronos forecasts candlestick series from its own PyTorch code; it is not a language model, has no chat template, tools or runtime support.",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/kronos"
    },
    {
      "slug": "laguna-m-1",
      "name": "Laguna M.1",
      "company": "Poolside, Inc. (named in Poolside's Terms of Use; the LICENSE.md copyright line reads Poolside)",
      "grade": "C",
      "fit": "no",
      "why": "Smallest build is the 139 GB NVFP4 for two Blackwell GPUs; nothing published fits a 16 GB to 128 GB machine, and there is no GGUF or Ollama entry.",
      "source": "https://huggingface.co/api/models/poolside/Laguna-M.1-NVFP4/tree/main?recursive=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/laguna-m-1"
    },
    {
      "slug": "lbs-27b",
      "name": "LBS-27B",
      "company": "inclusionAI, Ant Group (named in the site footer of inclusion-ai.org); the work is credited to authors from Wuhan University of Technology, CASIA, Zhejiang University and Ant Group",
      "grade": "C",
      "fit": "no",
      "why": "A GUI-grounding vision model that answers crop actions as plain text; the card says not to pass tools.",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/lbs-27b"
    },
    {
      "slug": "lfm2-5-2-6b",
      "name": "LFM2.5-2.6B",
      "company": "Liquid AI, Inc. (the licensor named in the LFM Open License)",
      "grade": "D",
      "fit": "no",
      "why": "A 2.6B on-device model that always thinks before answering; the card aims it at phones and laptops, not at a coding harness with a 10,000-token system prompt.",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/lfm2-5-2-6b"
    },
    {
      "slug": "minicpm-o-4-5",
      "name": "MiniCPM-o 4.5",
      "company": "OpenBMB (named as the MiniCPM-o team on the technical report; the GitHub LICENSE is the unfilled Apache template)",
      "grade": "C",
      "fit": "no",
      "why": "An omni-modal speech and vision model with a 40,960-token context; tool calling is not documented and the card aims it at live conversation, not a coding harness.",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/minicpm-o-4-5"
    },
    {
      "slug": "minicpm5-2b",
      "name": "MiniCPM5-2B",
      "company": "OpenBMB (named in the GitHub LICENSE's copyright line)",
      "grade": "C",
      "fit": "no",
      "why": "A 2.5B on-device model; the card aims it at edge use and documents only a thinking mode.",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/minicpm5-2b"
    },
    {
      "slug": "minimax-m3-1-flash",
      "name": "MiniMax-M3.1-Flash",
      "company": "MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)",
      "grade": "C",
      "fit": "no",
      "why": "Smallest GGUF is 129 GB (Unsloth UD-IQ1_M), above the 16 GB to 128 GB machines the use case names; the publisher's own build is 250 GB for 8 GPUs.",
      "source": "https://huggingface.co/api/models/unsloth/MiniMax-M3.1-Flash-GGUF/tree/main?recursive=true",
      "confidence": "low",
      "url": "https://www.anchorterminal.com/tools/minimax-m3-1-flash"
    },
    {
      "slug": "mistral-large-4",
      "name": "Mistral Large 4",
      "company": "Mistral AI (RCS Paris 952 418 325)",
      "grade": "C",
      "fit": "no",
      "why": "Smallest GGUF is 233 GB (Unsloth UD-IQ1_S), above the 16 GB to 128 GB machines the use case names.",
      "source": "https://huggingface.co/api/models/unsloth/Mistral-Large-4-1T-A52B-GGUF/tree/main?recursive=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/mistral-large-4"
    },
    {
      "slug": "nemotron-3-5-bolt-e4b-4gb",
      "name": "NVIDIA Nemotron 3.5 Bolt E4B 4GB",
      "company": "NVIDIA Corporation",
      "grade": "C",
      "fit": "no",
      "why": "NVIDIA aims it at voice and device assistants; 4B effective, 64K context and no coding benchmark on the card.",
      "source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB/raw/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/nemotron-3-5-bolt-e4b-4gb"
    },
    {
      "slug": "nemotron-3-nano-4b",
      "name": "Nemotron 3 Nano 4B",
      "company": "NVIDIA Corporation",
      "grade": "C",
      "fit": "no",
      "why": "NVIDIA aims it at game characters, voice assistants and IoT on edge devices, with no coding-agent benchmark on the card.",
      "source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16/raw/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/nemotron-3-nano-4b"
    },
    {
      "slug": "nemotron-3-nano-omni-30b-a3b-reasoning",
      "name": "Nemotron 3 Nano Omni 30B A3B Reasoning",
      "company": "NVIDIA Corporation",
      "grade": "C",
      "fit": "no",
      "why": "The card aims it at video, speech, documents and GUI agents, not coding, and every documented runtime command passes trust_remote_code.",
      "source": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16/raw/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/nemotron-3-nano-omni-30b-a3b-reasoning"
    },
    {
      "slug": "nemotron-3-ultra-550b-a55b",
      "name": "Nemotron 3 Ultra 550B A55B",
      "company": "NVIDIA Corporation",
      "grade": "B",
      "fit": "no",
      "why": "Smallest GGUF is 188 GB (Unsloth UD-IQ1_M), above the 16 GB to 128 GB machines the use case names.",
      "source": "https://huggingface.co/api/models/unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF/tree/main?recursive=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/nemotron-3-ultra-550b-a55b"
    },
    {
      "slug": "olmo-3-7b-think",
      "name": "Olmo 3 7B Think",
      "company": "The Allen Institute for Artificial Intelligence (Ai2), a 501(c)(3) nonprofit, per its Responsible Use Guidelines",
      "grade": "D",
      "fit": "no",
      "why": "The template reads a custom functions field rather than the OpenAI tools list and defaults to 'You do not currently have access to any functions'; no tool parser is named, thinking is always on and context is 64K.",
      "source": "https://huggingface.co/allenai/Olmo-3-7B-Think/raw/main/chat_template.jinja",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/olmo-3-7b-think"
    },
    {
      "slug": "ornith-1-5-397b",
      "name": "Ornith-1.5-397B",
      "company": "Ornith",
      "grade": "C",
      "fit": "no",
      "why": "Smallest build is the 244 GB Q4_K_M GGUF; nothing fits a 128 GB machine.",
      "source": "https://huggingface.co/api/models/ornith-ai/Ornith-1.5-397B-GGUF/tree/main?recursive=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/ornith-1-5-397b"
    },
    {
      "slug": "qwen3-5-0-8b",
      "name": "Qwen3.5-0.8B",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "D",
      "fit": "no",
      "why": "The card itself scopes this size to prototyping, fine-tuning and research; not aimed at driving a coding harness.",
      "source": "https://huggingface.co/Qwen/Qwen3.5-0.8B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-5-0-8b"
    },
    {
      "slug": "qwen3-5-2b",
      "name": "Qwen3.5-2B",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "D",
      "fit": "no",
      "why": "The card itself scopes this size to prototyping, fine-tuning and research; not aimed at driving a coding harness.",
      "source": "https://huggingface.co/Qwen/Qwen3.5-2B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-5-2b"
    },
    {
      "slug": "qwen3-5-397b-a17b",
      "name": "Qwen3.5-397B-A17B",
      "company": "Alibaba Cloud (named in the licence's copyright line)",
      "grade": "B",
      "fit": "no",
      "why": "Smallest build is Unsloth's 107 GB UD-IQ1_M; nothing fits the 16 to 128 GB machines the use case states at 4 bits.",
      "source": "https://huggingface.co/api/models/unsloth/Qwen3.5-397B-A17B-GGUF?blobs=true",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-5-397b-a17b"
    },
    {
      "slug": "qwen3-omni-30b-a3b-instruct",
      "name": "Qwen3-Omni-30B-A3B-Instruct",
      "company": "Alibaba Cloud (carried from the 20 October 2026 pilot; this model's files name no copyright holder)",
      "grade": "D",
      "fit": "no",
      "why": "An omni-modal chat model with a 65,536-token context; the card aims it at audio, vision and speech, not coding harnesses.",
      "source": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-omni-30b-a3b-instruct"
    },
    {
      "slug": "qwen3-vl-2b-instruct",
      "name": "Qwen3-VL-2B-Instruct",
      "company": "Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder)",
      "grade": "C",
      "fit": "no",
      "why": "A vision-language instruct model with Hermes-style JSON tool calls and no parser named; the card aims it at visual agents and OCR, not coding harnesses.",
      "source": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-vl-2b-instruct"
    },
    {
      "slug": "qwen3-vl-30b-a3b-instruct",
      "name": "Qwen3-VL-30B-A3B-Instruct",
      "company": "Alibaba Cloud (per the pilot's RDAP check of qwen.ai; this model's files name no copyright holder)",
      "grade": "C",
      "fit": "no",
      "why": "A vision-language instruct model with Hermes-style JSON tool calls and no parser named; the card aims it at visual agents and OCR, not coding harnesses.",
      "source": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-vl-30b-a3b-instruct"
    },
    {
      "slug": "qwen3-vl-32b-instruct",
      "name": "Qwen3-VL-32B-Instruct",
      "company": "Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder)",
      "grade": "C",
      "fit": "no",
      "why": "A vision-language instruct model with Hermes-style JSON tool calls and no parser named; the card aims it at visual agents and OCR, not coding harnesses.",
      "source": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-vl-32b-instruct"
    },
    {
      "slug": "qwen3-vl-4b-instruct",
      "name": "Qwen3-VL-4B-Instruct",
      "company": "Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder)",
      "grade": "D",
      "fit": "no",
      "why": "A vision-language instruct model with Hermes-style JSON tool calls and no parser named; the card aims it at visual agents and OCR, not coding harnesses.",
      "source": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-vl-4b-instruct"
    },
    {
      "slug": "qwen3-vl-8b-instruct",
      "name": "Qwen3-VL-8B-Instruct",
      "company": "Alibaba Cloud (per the pilot's RDAP check of qwen.ai; this model's files name no copyright holder)",
      "grade": "C",
      "fit": "no",
      "why": "A vision-language instruct model with Hermes-style JSON tool calls and no parser named; the card aims it at visual agents and OCR, not coding harnesses.",
      "source": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/qwen3-vl-8b-instruct"
    },
    {
      "slug": "ripple-65b-preview",
      "name": "Ripple-65B Preview",
      "company": "Ricursive Intelligence, Inc.",
      "grade": "D",
      "fit": "no",
      "why": "The documented path is Ricursive's SGLang launcher on two B200s with remote code; vLLM's stable list lacks the architecture and third-party GGUFs needed a patched llama.cpp, so no 16 to 128 GB machine runs it as shipped.",
      "source": "https://huggingface.co/ricursiveAI/Ripple-65B-Preview/raw/main/inference/server/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/ripple-65b-preview"
    },
    {
      "slug": "smollm3-3b",
      "name": "SmolLM3-3B",
      "company": "Hugging Face, Inc.",
      "grade": "C",
      "fit": "no",
      "why": "A 3B general model with a 65,536-token native context; the card aims it at on-device chat and tool use, not a coding harness sending 10K prompts.",
      "source": "https://huggingface.co/HuggingFaceTB/SmolLM3-3B",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/smollm3-3b"
    },
    {
      "slug": "unlimited-ocr-2",
      "name": "Baidu Unlimited-OCR-2",
      "company": "Baidu, Inc.",
      "grade": "C",
      "fit": "no",
      "why": "A document OCR model with no chat template or tool calling.",
      "source": "https://huggingface.co/baidu/Unlimited-OCR-2",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/unlimited-ocr-2"
    }
  ],
  "marks": "yes: the docs show it does the job as the success test says. caveat: it does the job with a limit an agent should know. no: it doesn't do this job. unchecked: we couldn't read it.",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-11-16",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.8",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "useCase": {
    "accounts": 27,
    "after": [
      "coding-agents.run-tests",
      "version-control.commit"
    ],
    "area": "models",
    "asks": [
      "run OpenCode against a local model",
      "replace Claude Code with something on my 3090",
      "which model for Pi on a Mac with 64 GB",
      "a coding agent that works offline"
    ],
    "before": [
      "coding-agents.harness-config"
    ],
    "blockers": [
      "tool calls emitted as prose the harness cannot execute",
      "KV cache rebuilt every turn on hybrid or Mamba models, so one turn takes minutes",
      "harness system prompt alone exceeds the context the memory allows",
      "smallest build does not fit the machine at 4 bits"
    ],
    "capabilities": [
      "model.open-weights",
      "model.tool-calling",
      "model.thinking-off"
    ],
    "category": "open-weights",
    "categoryName": "Open-weight models",
    "checked": "2026-11-12",
    "companies": {
      "fit": 2,
      "fitWithLimit": 34,
      "tagged": 90
    },
    "competition": "few",
    "fits": [
      "glm-4-7-flash",
      "qwen3-8-27b"
    ],
    "fitsWithLimit": [
      "agents-a1-4b",
      "apertus-v1-5-8b",
      "apodex-1-1-mini",
      "bigbang-v1",
      "darwin-36b-opus",
      "deepseek-v4-flash-0731",
      "gemma-4-12b-it",
      "gemma-4-26b-a4b-it",
      "gemma-4-31b-it",
      "gemma-4-e2b-it",
      "gemma-4-e4b-it",
      "glm-4-5-air",
      "glm-5-3-flash",
      "gpt-oss-120b",
      "gpt-oss-20b",
      "granite-4-2-3b",
      "hy3",
      "inkling-small",
      "kolibri-1",
      "laguna-s-2-1",
      "laguna-xs-2-1",
      "ling-3-0-tiny",
      "mimo-v2-6-distill-qwen-9b",
      "minimax-m2-5",
      "minimax-m2-7",
      "muse-glimmer-30b",
      "nanbeige-4-3-33b",
      "nemotron-3-5-lightning-30b-a3b",
      "nemotron-3-nano-30b-a3b",
      "nemotron-3-super-120b-a12b",
      "occamy-1-0",
      "ornith-1-5-35b-a3b",
      "ornith-1-5-9b",
      "qwen3-30b-a3b-instruct-2507",
      "qwen3-4b-instruct-2507",
      "qwen3-5-122b-a10b",
      "qwen3-5-4b",
      "qwen3-5-9b",
      "qwen3-6-27b",
      "qwen3-6-35b-a3b",
      "qwen3-8-flash-next",
      "qwen3-coder-30b-a3b-instruct",
      "qwen3-coder-next",
      "qwopus3-8-27b-flash-v2",
      "recursive-flint",
      "spark-x2-6-4b",
      "swift-1-5-qwen3-8-27b",
      "swift-1-5-qwen3-8-flash-next",
      "thinkingcap-qwen3-8-27b"
    ],
    "id": "open-weights.local-coding-agent",
    "listings": {
      "caveat": 49,
      "no": 35,
      "notRead": 28,
      "tagged": 161,
      "unchecked": 0,
      "yes": 2
    },
    "name": "Drive a coding agent harness from a local model",
    "status": "draft",
    "success": "Every tool call the model makes comes back in the tool_calls field, none as text. Measured: the model completes a scoped edit-and-test task with no edit loop past three retries, and decodes at least 20 tokens per second at 32K context on the stated machine.",
    "url": "https://www.anchorterminal.com/api/v1/use-cases/open-weights.local-coding-agent.json",
    "weight": 30,
    "weightBasis": "27 of 51 accounts; the dominant topic in every HN thread and forum read"
  }
}
