Niches · Open-weight models

Run a batch of files through the model overnight

Two or three companies' docs show a full fit.

2companies, full fit
43companies, fit with a limit
90companies carry the tag
12share of the category's work
7first-hand accounts

What counts as done

Output validates against the user's schema on the stated machine.

Measured, for our own tests: every item returns a non-empty output that validates against the schema, under 1 percent silent empties, and the whole batch finishes in one night on the stated machine (checker: 100 items, zero empties, all parse). No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: tag and summarise 1,000 markdown files to JSON, rename and sort my media folders, classify these rows without sending them out, overnight bulk job on my own GPU

Does the job 2

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

ListingWhy
DeepSeek-V4-Flash-0731 C
DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)
acc-41 ran 1,241 calls through DS4 at Q4 on Mac Studios with zero empty completions at 11.6 seconds mean; thinking_mode chat turns thinking off and deepseek-recipe renders json_schema. source
Gemma 4 31B IT BB
Google LLC (the registrant of gemma.ai; the card names Google DeepMind as author)
acc-10 runs structured automation with the 31B at Q4 and thinking off on a DGX Spark at about 20 tokens per second and finds the format correct almost every time; thinking is off by default. No JSON mode documented, so grammars come from the runtime. source

Does the job, with a limit 67

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.

ListingWhy
Agents-A1-4B C
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
Thinking can be switched off through the template and the 4B fits any machine, but structured output is undocumented and the template inserts a dated default system prompt when none is sent. source
Apodex-1.1-mini C
Apodex (the site's copyright line reads 2026 Apodex)
Thinking turns off through the template and the MoE runs fast at 3B active, but structured output is undocumented and the template keeps earlier reasoning by default (preserve_thinking). source
BigBang-v1 C
The Endless Frontier (endless-frontier)
Thinking can be disabled in the template and vLLM guided decoding works on the Qwen3.5 MoE class, but the card wants 128K of context and gives no structured-output guidance. source
Darwin-36B-Opus C
VIDRAFT Inc. (named on vidraft.net); the card's citation names FINAL-Bench and VIDRAFT_LAB
Thinking can be switched off in the template and the MoE is fast, but the card wants 5,120 tokens for think traces, documents no structured output and its stop tokens need <|im_end|> added. source
DeepSeek-V3.2 C
DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)
thinking_mode chat turns thinking off and vLLM 0.33.0 has deepseek_v32 tool and deepseek_v3 reasoning parsers, but the smallest build needs a 256 GB box and the repository has no Jinja template, only encoding_dsv32.py. source
DeepSeek V4.1 Flash C
DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)
enable_thinking false and reasoning_effort are in the repository's chat template, and deepseek-recipe renders json_schema response formats; but the machine must hold 233 GB or more and acc-41 ran the earlier V4 Flash, not V4.1. source
Gemma 4 12B IT BB
Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author)
Thinking off by default and a 7 GB build, but Google documents no JSON or schema mode, so you rely on runtime grammars; no bulk account for the 12B. source
Gemma 4 26B A4B IT BB
Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author)
acc-42 served live traffic on llama.cpp but moved the document batch to a hosted API because vLLM had no 4-bit MoE path; thinking is off by default and no JSON mode is documented. source
Gemma 4 E2B IT C
Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author)
Thinking off and tiny, but no JSON mode is documented and the card's figures put E2B weakest in the family; expect more invalid outputs. source
Gemma 4 E4B IT B
Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author)
Thinking off by default and 5 GB of weights, but no JSON mode is documented and no bulk account covers E4B. source
GLM-4.5-Air C
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)
enable_thinking false adds /nothink in the template and vLLM guided decoding applies, but Z.ai documents no structured output and the FP8 build needs two H100s; local batching means a third-party GGUF. source
GLM-4.7-Flash B
JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai)
Thinking turns off per request and vLLM 0.32 or later is a released path for continuous batching, but structured output is undocumented and context differs between config (202,752) and tokenizer (128,000). source
GLM-5.3 B
JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai). The licence names Z.AI
Thinking cannot be turned off, only lowered with reasoning_effort low (the template ignores enable_thinking); every item carries a think block and the smallest build needs 256 GB. source
GLM-5.3-Flash B
Z.AI Co., Ltd (named in the licence's copyright line)
Thinking is always on; reasoning_effort low shortens it but the template opens a think block on every item, and the smallest build needs about 100 GB. source
gpt-oss-120b BB
OpenAI (named as publisher on the card; the openai-api listing's terms name OpenAI OpCo, LLC)
Structured outputs are documented in the harmony guide and reasoning drops to low, but it never turns off, and acc-31 shows harmony tags leaking into grammar-constrained output on llama.cpp. source
gpt-oss-20b B
OpenAI (named in USAGE_POLICY and the model card; the LICENSE file's copyright line is the unfilled Apache template)
Structured outputs via a JSON Schema under Response Formats are documented and reasoning can be set low, but it never turns off and acc-31 shows harmony tags breaking grammars on llama-server between releases. source
Granite 4.2 3B C
International Business Machines Corporation (per IBM's naming guidance, as read for granite-guardian)
Thinking off and 7.3 GB BF16 make batches cheap and Ollama's page claims structured JSON output, but IBM's card documents no JSON mode, so schema validation falls to runtime guided decoding. source
Hy3 B
Tencent (named in the LICENSE file's copyright line)
reasoning_effort no_think is the default and the template enforces it, so thinking is off; the smallest build is 89 GB and structured output is undocumented. source
Inkling-Small B
Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary)
reasoning_effort none conditions the model away from thinking without forbidding it, structured output is undocumented, and the smallest build is 75 GB. source
InternVL3.5-GPT-OSS-20B-A4B-Preview D
Shanghai AI Laboratory (OpenGVLab), not named in the repository
vLLM 0.33.0 loads InternVLForConditionalGeneration and thinking is off unless the R1 system prompt is sent, but structured output is undocumented, the template uses markers the tokenizer lacks, and 42.5 GB of BF16 needs a 64 GB card. source
Kolibri-1 C
Aleph Alpha GmbH
enable_thinking false or reasoning_effort none turn thinking off and the card claims structured extraction, but it never shows how to request a schema and the FP8 build needs about 95 GB. source
Laguna M.1 C
Poolside, Inc. (named in Poolside's Terms of Use; the LICENSE.md copyright line reads Poolside)
Runs only on a multi-GPU Blackwell or Hopper server (NVFP4 139 GB, FP8 larger); thinking is off by default, vLLM poolside_v1 parsers named, but structured output is not documented. source
Laguna S 2.1 C
Poolside, Inc. (named in Poolside's Terms of Use)
Fits a 128 GB machine at Q4_K_M with enable_thinking false per request, but structured output is not documented and discussions report looping; smaller machines cannot run it. source
Laguna XS 2.1 C
Poolside, Inc. (named in Poolside's Terms of Use)
Fits 32 GB machines with thinking off by default, but structured output is undocumented and the card asks you to validate JSON arguments. source
LFM2.5-2.6B D
Liquid AI, Inc. (the licensor named in the LFM Open License)
Fits any machine and llama.cpp lists LFM2, but thinking is always on and structured output is undocumented. source
Ling-3.0-tiny C
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
Fits any GPU with thinking off, but structured output is undocumented and the documented vLLM route runs repository code. source
MiMo-V2.6 Distill Qwen 9B C
Xiaomi (the card and report name Xiaomi MiMo; no LICENSE file names a copyright holder)
Fits and switches thinking off, but structured output is undocumented and runaway generation is reported, so cap max_tokens. source
MiniCPM-o 4.5 C
OpenBMB (named as the MiniCPM-o team on the technical report; the GitHub LICENSE is the unfilled Apache template)
Fits any GPU and llama.cpp runs it, but structured output is undocumented, context is capped at 40,960 tokens and vLLM needs trust_remote_code. source
MiniCPM5-2B C
OpenBMB (named in the GitHub LICENSE's copyright line)
Fits anywhere on a plain Llama architecture, but thinking is always on and structured output is undocumented. source
MiniMax-M2.5 C
MiniMax AI (named in the licences; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)
A 128 GB box runs a 3-bit GGUF, but thinking cannot be turned off, structured output is undocumented, and smaller GPUs are out. source
MiniMax-M2.7 C
MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)
Runs on a 128 GB box at 3 bits for personal or research batches, but thinking spends tokens on every item, structured output is undocumented, and commercial batches need MiniMax's written authorisation. source
Muse Glimmer 30B BB
Meta (the verified Hugging Face organisation is "Meta Inc.", the card names Meta Superintelligence Lab, and the licence carries no copyright line)
Fits a 24 GB GPU, but structured output is undocumented, thinking is always on, and a request cut off mid-thought returns empty content in llama.cpp. source
Nanbeige4.3-33B C
BOSS Zhipin (Kanzhun Limited), unconfirmed
Thinking can be switched off and vLLM serves the plain Llama architecture with continuous batching, but structured output is undocumented and a 24 GB-plus GPU is needed. source
NVIDIA Nemotron 3.5 Bolt E4B 4GB C
NVIDIA Corporation
Thinking off, StructEval-T 80 at Q4_K_M and 6 GB, so any 8 GB GPU runs a night's batch; structured output is NVIDIA's benchmark, not a documented parameter, and 64K caps file size. source
Nemotron 3.5 Lightning 30B A3B B
NVIDIA Corporation
Thinking switches off and vLLM batches BF16 or NVFP4 on one GPU; structured output is a training stage on the card, not a documented parameter, and NVFP4 needs an NVIDIA GPU. source
Nemotron 3 Nano 30B A3B B
NVIDIA Corporation
Thinking switches off and vLLM batches the FP8 or NVFP4 build on one GPU, but structured output is undocumented and the vLLM path runs repository code. source
Nemotron 3 Nano 4B C
NVIDIA Corporation
Small and fast with thinking off, but structured output isn't documented, and English only. source
Nemotron 3 Nano Omni 30B A3B Reasoning C
NVIDIA Corporation
Instruct mode and a page-by-page PDF script on the card suit overnight document batches on a 32 GB GPU, but structured output is undocumented and vLLM loads repository code. source
Nemotron 3 Super 120B A12B B
NVIDIA Corporation
Thinking switches off or to low_effort per request and a 128 GB box holds the 4-bit build, but structured output is undocumented and decode is about 17 tok/s on Strix Halo. source
occamy-1.0 C
Accio (Accio-Lab)
BF16, FP8 and NVFP4 passed Accio's vLLM check covering JSON and tool calls, and thinking switches off; structured output is otherwise undocumented and GGUF input must be NFC-normalised. source
Olmo 3 7B Think D
The Allen Institute for Artificial Intelligence (Ai2), a 501(c)(3) nonprofit, per its Responsible Use Guidelines
4.5 GB in Ollama and Apache 2.0, so a night's batch fits any GPU, but every item spends a long think block and structured output is undocumented. source
Ornith-1.5-35B-A3B C
Ornith
Fits 24 to 32 GB with thinking off, but structured output is undocumented and looping is reported. source
Ornith-1.5-9B C
Ornith
Fits anywhere with thinking off, but structured output is undocumented. source
Qwen3-30B-A3B-Instruct-2507 C
Alibaba Cloud (named in the licence's copyright line)
Non-thinking, so no reasoning eats the budget, and vLLM batches the BF16 or FP8 build; structured output is undocumented, so JSON validity rests on the runtime's guided decoding. source
Qwen3-4B-Instruct-2507 C
Alibaba Cloud (named in the licence's copyright line)
Non-thinking, 9 GB of KV at 64K, fits any 16 GB machine; structured output is undocumented, so JSON validity rests on the runtime's guided decoding. source
Qwen3.5-0.8B D
Alibaba Cloud (named in the licence's copyright line)
Fits anywhere (1.7 GB BF16, 0.8 GB KV at 64K) with thinking off, but structured output is undocumented and the card warns of thinking loops at this size. source
Qwen3.5-122B-A10B B
Alibaba Cloud (named in the licence's copyright line)
acc-41 ran 437 files through qwen3.5:122b in Ollama overnight but saw 18 to 20 percent silent empties at 16K context and 150 to 200 s per call; structured output is undocumented. source
Qwen3.5-2B D
Alibaba Cloud (named in the licence's copyright line)
Fits anywhere (4.5 GB BF16, 0.8 GB KV at 64K) with thinking off, but structured output is undocumented and the card warns of thinking loops at this size. source
Qwen3.5-4B C
Alibaba Cloud (named in the licence's copyright line)
Fits any machine with thinking off, but structured output is undocumented; JSON validity rests on vLLM or llama.cpp grammar constraints, untested with the template. source
Qwen3.5-9B C
Alibaba Cloud (named in the licence's copyright line)
Fits a 16 GB machine with thinking off and 2 GB of KV at 64K, but structured output is undocumented and no bulk account exists for this size. source
Qwen3.6-27B B
Alibaba Cloud (named in the licence's copyright line)
acc-13 runs it for overnight bulk jobs on 32 GB and acc-50 for summarisation; thinking off per request, but structured output is undocumented beyond runtime grammars. source
Qwen3.6-35B-A3B B
Alibaba Cloud (named in the licence's copyright line)
acc-14 ran bulk media renaming and library comparison on it; thinking off per request, but structured output is undocumented and vLLM has no 4-bit MoE path named on the card. source
Qwen3.8-27B B
Alibaba Cloud (named in the licence's copyright line)
Thinking off per request and reasoning_effort low, fits 24 GB with 4 GB of KV at 64K; structured output is undocumented beyond runtime grammars, and no bulk account exists. source
Qwen3.8-Flash-Next C
Alibaba Cloud (Singapore) Private Limited, the registrant of qwen.ai. The licence names Qwen as copyright holder
6B active decodes fast and thinking is off per request, but only a 128 GB machine holds it and structured output is undocumented; the Qwen Community License bars model-as-a-service use. source
Qwen3-Coder-30B-A3B-Instruct C
Alibaba Cloud (named in the licence's copyright line)
Non-thinking and fits 24 GB with 6 GB of KV at 64K, but it is tuned for code and structured output is undocumented beyond runtime grammars. source
Qwen3-Coder-Next C
Alibaba Cloud (the Qwen team's employer; no LICENSE file names a copyright holder for this model)
Non-thinking with 3B active and 6 GB of KV at 64K, but it needs a 64 GB machine, is tuned for code, and structured output is undocumented. source
Qwen3-VL-2B-Instruct C
Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder)
Non-thinking and fits anything; acc-43 looped a directory of scans through the 8B overnight, but structured output is undocumented beyond runtime grammars. source
Qwen3-VL-30B-A3B-Instruct C
Alibaba Cloud (per the pilot's RDAP check of qwen.ai; this model's files name no copyright holder)
Non-thinking and fits 32 GB with 6 GB of KV at 64K; acc-43 looped a directory of scans through the 8B overnight, but structured output is undocumented beyond runtime grammars. source
Qwen3-VL-32B-Instruct C
Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder)
Non-thinking and fits 32 GB only at short context, 16 GB of KV at 64K; acc-43 looped a directory of scans through the 8B overnight, but structured output is undocumented beyond runtime grammars. source
Qwen3-VL-4B-Instruct D
Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder)
Non-thinking and fits a 16 GB laptop with 9 GB of KV at 64K; acc-43 looped a directory of scans through the 8B overnight, but structured output is undocumented beyond runtime grammars. source
Qwen3-VL-8B-Instruct C
Alibaba Cloud (per the pilot's RDAP check of qwen.ai; this model's files name no copyright holder)
Non-thinking and fits 16 GB with 9 GB of KV at 64K; acc-43 looped a directory of scans through the 8B overnight, but structured output is undocumented beyond runtime grammars. source
Qwopus3.8-27B-Flash-V2 C
Jackrong, an individual Hugging Face account
enable_thinking false and reasoning_effort work in the inherited template, 17 GB fits 24 GB, but structured output is undocumented and the card warns of reasoning drift and loops. source
SmolLM3-3B C
Hugging Face, Inc.
Thinking off per request, 4.5 GB of KV at 64K, fits anything; structured output is undocumented beyond runtime grammars. source
Spark-X2.6-4B C
XHToken (the licence's copyright line); the Hugging Face organisation's name is SparkLLM
Thinking off per request, 2.3 GB of KV at 64K, fits anything; structured output is undocumented and vLLM needs SparkLLM's remote code. source
Swift 1.5 Qwen3.8-27B C
UkisAI
enable_thinking false and reasoning_effort per request on the base template, 17 GB fits 24 GB; structured output undocumented and the revenue-capped licence applies to the batch's business. source
ThinkingCap (Qwen3.8 27B) C
BottleCap AI (named in the licence's copyright line)
enable_thinking false and reasoning_effort per request in the inherited template and 17 GB fits 24 GB, but structured output is undocumented and the licence limits who may run the batch. source
Baidu Unlimited-OCR-2 C
Baidu, Inc.
Built to parse hundreds of pages per request, but output is layout blocks for a converter, not a JSON schema, and the documented path is remote code or Baidu's custom SGLang wheel on NVIDIA. source

Doesn't do this job 17

Listings in the category that do something else, with the reason.

ListingWhy
Apertus v1.5 8B C
ETH Zurich and EPFL (the Swiss National AI Institute, named in the Acceptable Use Policy)
No released runtime loads the model, so a continuous-batching run depends on the publisher's vLLM fork image; structured output is undocumented. source
Carbon-X C
Hugging Face (the HuggingFaceBio organisation, named Hugging Face Biology Research on the Hub)
Trained to answer about DNA sequences, not to tag or classify documents; the card's batch guidance is about sequence files. source
Cosmos-Reason2-2B C
NVIDIA Corporation
Built for video and image reasoning, not document batches; NVIDIA's JSON comes from prompt files for boxes and trajectories. source
DiffusionGemma 26B A4B IT D
Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author)
SGLang is the documented server and the listing's notes record it rejecting JSON Schema for this model; Google labels the model experimental and below Gemma 4 on most tasks. source
Inkling B
Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary)
reasoning_effort none exists, but no machine under 290 GB loads it and structured output is undocumented. source
Kimi K3 B
Moonshot AI (the licence names Moonshot AI; the Kimi Platform terms name Moonshot AI PTE. LTD.)
Thinking is always on with reasoning_effort low as the floor, and no machine under about 600 GB loads it. source
Kronos C
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
Takes numeric bar series only; cannot read documents or emit JSON against a schema.
LBS-27B C
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
Aims at screenshot grounding, not document extraction; no JSON schema path.
MiniMax-M3.1-Flash C
MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)
Needs a 192 GB-plus box or a multi-GPU node; structured output is undocumented and thinking cannot be turned off, only set to low effort. source
Mistral Large 4 C
Mistral AI (RCS Paris 952 418 325)
json_schema with strict true is documented, but the hardware is an 8-GPU node or a 256 GB-plus box at 1 to 2 bits, not one overnight machine of the kind the use case names. source
Nemotron 3 Ultra 550B A55B B
NVIDIA Corporation
NVIDIA's smallest build needs four B200-class GPUs and the GGUFs a 256 GB-plus box; not a one-machine overnight job of the kind the use case names, and structured output is undocumented. source
Ornith-1.5-397B C
Ornith
Needs about 300 GB of memory or an 8-GPU server; not a single-GPU job. source
Qwen3.5-397B-A17B B
Alibaba Cloud (named in the licence's copyright line)
Needs an 8-GPU server or a 256 GB box; not a single-machine overnight job in the stated sense, and structured output is undocumented. source
Qwen3-Omni-30B-A3B-Instruct D
Alibaba Cloud (carried from the 20 October 2026 pilot; this model's files name no copyright holder)
Aimed at audio and video understanding; structured output is undocumented and the use case's document batches are not its job. source
Recursive Flint C
Recursive Superintelligence, Inc. (the Hugging Face organisation's full name)
The SGLang kit documents response_format json_schema, but enable_thinking has no effect, so every item carries a reasoning block and the use case's thinking-off capability is missing. source
Ripple-65B Preview D
Ricursive Intelligence, Inc.
Always-on thinking fails the thinking-off capability, and the only documented server is a two-B200 launcher. source
Swift 1.5 Qwen3.8-Flash-Next C
UkisAI
Needs a 128 GB machine at 3 bits, and both the Swift revenue cap and the Qwen Community License's model-as-a-service clause sit on top of undocumented structured output. source

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

What stops agents finishing

  • 18 to 20 percent silent empty completions at 16K context
  • vLLM has no 4-bit MoE path for the model, so continuous batching is unavailable
  • files over the usable context need chunking the user did not plan for

Around it

Before: document-extraction.pdf-to-text, file-storage.list. After: vector-databases.upsert, spreadsheets.write-rows.

Where this comes from

Weight 12 of 100 in Open-weight models: 7 of 51 accounts; the second most common non-coding job in the HN threads

The same as data: /api/v1/use-cases/open-weights.bulk-extraction.json, or the MCP tool get_use_case with id open-weights.bulk-extraction. Other jobs in this category: /api/v1/niches?category=open-weights.

Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.