# Run a frontier-class MoE on 128 GB to 512 GB of unified memory

> Run a frontier-class MoE on 128 GB to 512 GB of unified memory: 3 companies fully do it, 14 with a limit, 90 carry the tag. Every listing in Open-weight models with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/open-weights.big-model-on-big-box/
- Markdown: https://www.anchorterminal.com/niches/open-weights.big-model-on-big-box/index.md (~4,850 tokens)
- Slim: https://www.anchorterminal.com/niches/open-weights.big-model-on-big-box/index.min.md (~780 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/open-weights.big-model-on-big-box/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Two or three companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 3 | 14 | 90 | 10 | 6 |

## What counts as done

The model loads and holds a 64K context in memory without swapping.

Measured, for our own tests: decode at 10 tokens per second or better, prompt processing above 100 tokens per second, and answers in the prompt's language.

How agents ask for it: run DeepSeek V4 Flash on my Mac Studio; what fits a 192 GB Framework Desktop; GLM-5.3 at Q2 as a reviewer; MiniMax at 100K context on Strix Halo

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **DeepSeek-V3.2 (DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.))**: Unsloth GGUFs from 161 GB (TQ1_0) to 217 GB (IQ2_XXS) load on 256 to 512 GB under llama.cpp's deepseek2 architecture; MLA keeps the 64K cache near 5 GB. No account times decode or prompt processing.
- **gpt-oss-120b (OpenAI (named as publisher on the card; the openai-api listing's terms name OpenAI OpCo, LLC))**: 117B MoE at 63 GB MXFP4 with 5.1B active; acc-21 and acc-40 ran it on 128 GB Strix Halo machines alongside other models, Ollama's gpt-oss:120b is 65 GB, llama.cpp has gpt-oss and 64K of cache is about 5 GB.
- **Laguna S 2.1 (Poolside, Inc. (named in Poolside's Terms of Use))**: 96 GB Q4_K_M plus about 12 GiB KV at 64K holds in 128 GB; an account measured 186 tokens per second prompt processing and 27.8 decode at UD-Q4_K_XL, Ollama ships the tag, and llama.cpp lists the laguna architecture.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **DeepSeek V4.1 Flash (DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.))**: acc-22 runs it at Q2 on a 192 GB Framework Desktop near 14 tokens per second; llama.cpp has deepseek41 and Unsloth ships IQ1_S (233 GB) to Q4_K_XL (501 GB). Prompt-processing speed is unreported and acc-40 saw Chinese replies from V4 Flash at IQ3.
- **DeepSeek-V4-Flash-0731 (DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.))**: Loads on 128 GB at IQ2 (acc-08, 30 to 35 tokens per second) and Unsloth recommends IQ3_XXS at 103 GB with 110 GB of RAM; llama.cpp has deepseek4. acc-40 saw Chinese answers to English prompts at IQ3_XXS.
- **GLM-4.5-Air (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))**: A 106B MoE, below the 500B class, but Unsloth's Q4_K_XL (68 GB) plus a 12 GB bf16 cache at 64K fits 128 GB with 12B active; glm4moe is in llama.cpp. No account times it.
- **GLM-5.3 (JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai). The licence names Z.AI)**: Unsloth's UD-IQ2_M (239 GB) targets 256 GB boxes and llama.cpp has glm-dsa; 40B active at 2 bits has no first-hand speed report, and acc-22 found the Flash cousin too slow for interactive use at Q2.
- **GLM-5.3-Flash (Z.AI Co., Ltd (named in the licence's copyright line))**: 320B MoE with 18B active; Unsloth puts 1-bit at 100 GB and 3-bit at 128 to 150 GB, llama.cpp has glm5-next, and acc-22 reached a usable reviewer at Q2 on 192 GB but calls it slow for interactive work.
- **Hy3 (Tencent (named in the LICENSE file's copyright line))**: 295B MoE with 21B active; AngelSlim ships IQ1_M (89 GB), Q2_K_XL (100 GB), Q3_K_M (135 GB) and Q4_K_M (182 GB) with a llama-server line at 64K and a q8 cache. No first-hand speed or language report.
- **Inkling (Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary))**: 975B MoE; Unsloth's 1-bit needs about 290 GB and 2-bit 325 GB, so only a 512 GB box qualifies. The inkling architecture is not in llama.cpp master's list, so the GGUF path rests on Unsloth's statement that it runs.
- **Inkling-Small (Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary))**: 276B MoE with 12B active; Unsloth's table gives 89 GB at 2-bit and 132 to 170 GB at 4-bit, inside the 128 to 256 GB band, but llama.cpp master lacks the inkling architecture and no account times it.
- **Kolibri-1 (Aleph Alpha GmbH)**: A 78B MoE with 3.5B active, under the 500B class; FP8 (78.9 GB) or the Q4 GGUF (49.8 GB) fits 128 GB with a 6.7 GB cache at 64K (40 of 50 layers sliding window). No first-hand speed report.
- **MiniMax-M2.5 (MiniMax AI (named in the licences; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore))**: Unsloth GGUFs from 74 GB (IQ2_XXS) to 131 GB (UD-Q4_K_XL) fit 128 GB to 256 GB, llama.cpp lists MINIMAXM2 and LM Studio states 121 GB minimum, but no account gives decode or prompt-processing speed and every reply spends tokens on thinking.
- **MiniMax-M2.7 (MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore))**: Unsloth GGUFs from 61 GB to 141 GB (UD-Q4_K_XL) fit 128 GB to 256 GB boxes; a second-hand Strix Halo account reached 100K at UD-IQ3_XXS only with mmap off and swap added, and prompt processing was the painful part.
- **MiniMax-M3.1-Flash (MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore))**: Unsloth GGUFs of 135 GB to 249 GB fit 192 GB to 512 GB boxes with a llama.cpp build carrying MiniMax-M3.1, plus a DSpark draft; nothing fits 128 GB and no account gives speeds.
- **Mistral Large 4 (Mistral AI (RCS Paris 952 418 325))**: ggml-org and Unsloth publish GGUFs with mmproj, but only 1 to 2 bit builds (233 GB to 378 GB) fit a 256 GB or 512 GB box, nothing fits 128 GB to 192 GB, and no account gives speeds.
- **Nemotron 3 Super 120B A12B (NVIDIA Corporation)**: A 120B rather than 500B-class model, but it is what fits: Unsloth 4-bit needs 64 GB to 72 GB, SEP measured 17 tok/s on a 128 GB Strix Halo after raising the GPU memory limit; prompt-processing speed not given.
- **Nemotron 3 Ultra 550B A55B (NVIDIA Corporation)**: Unsloth GGUFs from 188 GB to 360 GB (UD-Q4_K_XL) and a 347 GB MLX 4-bit build fit 256 GB to 512 GB boxes; nothing fits 128 GB or 192 GB, and no account gives speeds for 55B active parameters.
- **Ornith-1.5-397B (Ornith)**: Q4_K_M 244 GB plus 1.9 GiB KV fits a 512 GB box and Ollama lists a 397b tag at 242 GB, but Ornith ships no 2 or 3 bit build, no account gives speeds and the licence exists only as a tag.
- **Qwen3.5-122B-A10B (Alibaba Cloud (named in the licence's copyright line))**: Not 500B class, but the 128 GB owners' pick: Ollama 81 GB Q4 or Unsloth UD-Q4_K_XL 77 GB with 1.5 GB of KV at 64K; acc-49 found it accurate but slow.
- **Qwen3.5-397B-A17B (Alibaba Cloud (named in the licence's copyright line))**: The family's 500B-class MoE: Unsloth UD-IQ2_M 123 GB or UD-Q3_K_XL 179 GB with 1.9 GB of KV at 64K fit 192 to 512 GB, but no account runs it and Ollama has no tag.
- **Qwen3.8-Flash-Next (Alibaba Cloud (Singapore) Private Limited, the registrant of qwen.ai. The licence names Qwen as copyright holder)**: 125B MoE for 128 GB boxes: Ollama 105 GB MLX on a Mac, Unsloth UD-IQ4_XS 93.7 GB elsewhere, 1.5 GB of KV at 64K; vLLM's stable list lacks the qwen4_exp architecture.
- **Swift 1.5 Qwen3.8-Flash-Next (UkisAI)**: 180B-file MoE for 128 GB boxes only through the 66.6 to 84.7 GB GSQ-RCO quants with 1.5 GB of KV at 64K; the qwen4_exp architecture is missing from vLLM's stable list and no account runs it.

- not this job: Agents-A1-4B (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): A 4.5B dense model, not a 500B-class mixture of experts; the use case does not apply.
- not this job: Apertus v1.5 8B (ETH Zurich and EPFL (the Swiss National AI Institute, named in the Acceptable Use Policy)): An 8.9B dense model, not a 500B-class mixture of experts.
- not this job: Apodex-1.1-mini (Apodex (the site's copyright line reads 2026 Apodex)): A 36B mixture of experts that fits 32 GB; not the 500B-class checkpoint this use case is about.
- not this job: BigBang-v1 (The Endless Frontier (endless-frontier)): A 36B mixture of experts, not a 500B-class checkpoint.
- not this job: Carbon-X (Hugging Face (the HuggingFaceBio organisation, named Hugging Face Biology Research on the Hub)): 14.8B dense, not a 500B-class checkpoint.
- not this job: Cosmos-Reason2-2B (NVIDIA Corporation): A 2.4B model.
- not this job: Darwin-36B-Opus (VIDRAFT Inc. (named on vidraft.net); the card's citation names FINAL-Bench and VIDRAFT_LAB): A 35B mixture of experts that fits 32 GB.
- not this job: DiffusionGemma 26B A4B IT (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author)): A 25B mixture of experts.
- not this job: Gemma 4 12B IT (Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author)): A 12B dense model.
- not this job: Gemma 4 26B A4B IT (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author)): A 26B mixture of experts that fits 24 GB.
- not this job: Gemma 4 31B IT (Google LLC (the registrant of gemma.ai; the card names Google DeepMind as author)): A 31B dense model.
- not this job: Gemma 4 E2B IT (Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author)): A 2B-effective model.
- not this job: Gemma 4 E4B IT (Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author)): A 4B-effective model.
- not this job: GLM-4.7-Flash (JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai)): A 30B mixture of experts that fits 24 GB.
- not this job: gpt-oss-20b (OpenAI (named in USAGE_POLICY and the model card; the LICENSE file's copyright line is the unfilled Apache template)): A 21B mixture of experts that fits 16 GB.
- not this job: Granite 4.2 3B (International Business Machines Corporation (per IBM's naming guidance, as read for granite-guardian)): A 3.7B dense model.
- not this job: InternVL3.5-GPT-OSS-20B-A4B-Preview (Shanghai AI Laboratory (OpenGVLab), not named in the repository): A 21B mixture of experts.
- not this job: Kimi K3 (Moonshot AI (the licence names Moonshot AI; the Kimi Platform terms name Moonshot AI PTE. LTD.)): Unsloth's table puts the smallest dynamic 1-bit at 610 GB of memory and the 466 GB UD-Q1_0 has no quality figure; a 512 GB machine falls short. llama.cpp has kimi-k3.
- not this job: Kronos (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): Largest size is Kronos-base, 0.43 GB in FP32; not a mixture-of-experts language model.
- not this job: Laguna M.1 (Poolside, Inc. (named in Poolside's Terms of Use; the LICENSE.md copyright line reads Poolside)): No GGUF or MLX build from Poolside or third parties was found, and the 139 GB NVFP4 build needs NVIDIA Blackwell GPUs rather than unified memory.
- not this job: Laguna XS 2.1 (Poolside, Inc. (named in Poolside's Terms of Use)): 33B model; does not aim at the 500B-class use case.
- not this job: LBS-27B (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): 27B dense vision model; not the use case.
- not this job: LFM2.5-2.6B (Liquid AI, Inc. (the licensor named in the LFM Open License)): 2.6B model.
- not this job: Ling-3.0-tiny (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): 7.9B model.
- not this job: MiMo-V2.6 Distill Qwen 9B (Xiaomi (the card and report name Xiaomi MiMo; no LICENSE file names a copyright holder)): 9B model.
- not this job: MiniCPM-o 4.5 (OpenBMB (named as the MiniCPM-o team on the technical report; the GitHub LICENSE is the unfilled Apache template)): 9.4B model.
- not this job: MiniCPM5-2B (OpenBMB (named in the GitHub LICENSE's copyright line)): 2.5B model.
- not this job: Muse Glimmer 30B (Meta (the verified Hugging Face organisation is "Meta Inc.", the card names Meta Superintelligence Lab, and the licence carries no copyright line)): A 30B dense model, not a 500B-class mixture of experts.
- not this job: Nanbeige4.3-33B (BOSS Zhipin (Kanzhun Limited), unconfirmed): A 33B dense model, not a 500B-class mixture of experts.
- not this job: NVIDIA Nemotron 3.5 Bolt E4B 4GB (NVIDIA Corporation): An 8B (4B effective) model, not a 500B-class mixture of experts.
- not this job: Nemotron 3.5 Lightning 30B A3B (NVIDIA Corporation): A 30B mixture of experts, not a 500B-class model.
- not this job: Nemotron 3 Nano 30B A3B (NVIDIA Corporation): A 30B mixture of experts, not a 500B-class model.
- not this job: Nemotron 3 Nano 4B (NVIDIA Corporation): A 4B model, not a 500B-class mixture of experts.
- not this job: Nemotron 3 Nano Omni 30B A3B Reasoning (NVIDIA Corporation): A 31B mixture of experts, not a 500B-class model.
- not this job: occamy-1.0 (Accio (Accio-Lab)): A 35B mixture of experts, not a 500B-class model.
- not this job: Olmo 3 7B Think (The Allen Institute for Artificial Intelligence (Ai2), a 501(c)(3) nonprofit, per its Responsible Use Guidelines): A 7B dense model, not a 500B-class mixture of experts.
- not this job: Ornith-1.5-35B-A3B (Ornith): 35B model.
- not this job: Ornith-1.5-9B (Ornith): 9B model.
- not this job: Qwen3-30B-A3B-Instruct-2507 (Alibaba Cloud (named in the licence's copyright line)): A 30B mixture of experts, not a 500B-class model.
- not this job: Qwen3-4B-Instruct-2507 (Alibaba Cloud (named in the licence's copyright line)): A 4B dense model, not a 500B-class mixture of experts.
- not this job: Qwen3.5-0.8B (Alibaba Cloud (named in the licence's copyright line)): A small dense model, not a 500B-class mixture of experts.
- not this job: Qwen3.5-2B (Alibaba Cloud (named in the licence's copyright line)): A small dense model, not a 500B-class mixture of experts.
- not this job: Qwen3.5-4B (Alibaba Cloud (named in the licence's copyright line)): A 4B dense model, not a 500B-class mixture of experts.
- not this job: Qwen3.5-9B (Alibaba Cloud (named in the licence's copyright line)): A 9B dense model, not a 500B-class mixture of experts.
- not this job: Qwen3.6-27B (Alibaba Cloud (named in the licence's copyright line)): A 27B dense model, not a 500B-class mixture of experts.
- not this job: Qwen3.6-35B-A3B (Alibaba Cloud (named in the licence's copyright line)): A 35B mixture of experts with 3B active, not 500B class.
- not this job: Qwen3.8-27B (Alibaba Cloud (named in the licence's copyright line)): A 27B dense model, not a 500B-class mixture of experts.
- not this job: Qwen3-Coder-30B-A3B-Instruct (Alibaba Cloud (named in the licence's copyright line)): A 30B mixture of experts, not 500B class.
- not this job: Qwen3-Coder-Next (Alibaba Cloud (the Qwen team's employer; no LICENSE file names a copyright holder for this model)): An 80B mixture of experts that fits 64 GB; not the 500B-class job this use case describes.
- not this job: Qwen3-Omni-30B-A3B-Instruct (Alibaba Cloud (carried from the 20 October 2026 pilot; this model's files name no copyright holder)): A 30B mixture of experts, not 500B class.
- not this job: Qwen3-VL-2B-Instruct (Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder)): Not a 500B-class mixture of experts.
- not this job: Qwen3-VL-30B-A3B-Instruct (Alibaba Cloud (per the pilot's RDAP check of qwen.ai; this model's files name no copyright holder)): Not a 500B-class mixture of experts.
- not this job: Qwen3-VL-32B-Instruct (Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder)): Not a 500B-class mixture of experts.
- not this job: Qwen3-VL-4B-Instruct (Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder)): Not a 500B-class mixture of experts.
- not this job: Qwen3-VL-8B-Instruct (Alibaba Cloud (per the pilot's RDAP check of qwen.ai; this model's files name no copyright holder)): Not a 500B-class mixture of experts.
- not this job: Qwopus3.8-27B-Flash-V2 (Jackrong, an individual Hugging Face account): A 27B dense fine-tune, not a 500B-class mixture of experts.
- not this job: Recursive Flint (Recursive Superintelligence, Inc. (the Hugging Face organisation's full name)): A 40B mixture of experts with 2.5B active, not 500B class.
- not this job: Ripple-65B Preview (Ricursive Intelligence, Inc.): A 65B dense model, not a 500B-class mixture of experts, and no publisher GGUF.
- not this job: SmolLM3-3B (Hugging Face, Inc.): A 3B dense model, not a 500B-class mixture of experts.
- not this job: Spark-X2.6-4B (XHToken (the licence's copyright line); the Hugging Face organisation's name is SparkLLM): A 4B dense model, not a 500B-class mixture of experts.
- not this job: Swift 1.5 Qwen3.8-27B (UkisAI): A 27B dense fine-tune, not a 500B-class mixture of experts.
- not this job: ThinkingCap (Qwen3.8 27B) (BottleCap AI (named in the licence's copyright line)): A 27B dense fine-tune, not a 500B-class mixture of experts.
- not this job: Baidu Unlimited-OCR-2 (Baidu, Inc.): A 3.3B OCR model, not a 500B-class mixture of experts.

28 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- smallest build larger than the machine at any precision
- prompt processing so slow that one agent turn takes minutes
- model answers in Chinese to English prompts at IQ3
- runtime support only in a nightly or Docker tag

## Around it

Before: open-weights.local-coding-agent. After: coding-agents.plan-review.

## Where this comes from

Weight 10 of 100 in Open-weight models: 6 of 51 accounts, all from owners of 128 GB or larger machines

As data: https://www.anchorterminal.com/api/v1/use-cases/open-weights.big-model-on-big-box.json, or the MCP tool get_use_case with id open-weights.big-model-on-big-box. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
