Niches · Open-weight models

Read scans, screenshots and PDF pages locally

Four or more companies' docs show a full fit.

8companies, full fit
14companies, fit with a limit
90companies carry the tag
8share of the category's work
4first-hand accounts

What counts as done

Text or JSON out for every page.

Measured, for our own tests: word error rate better than Tesseract on the user's sample, and under 60 seconds per page on the stated machine. No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: OCR these typewritten scans, pull fields from receipts into JSON, read this screenshot, unify thousands of medical record pages

Does the job 10

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

ListingWhy
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)
Google ships mmproj-gemma-4-12b-it-qat-q4_0.gguf (0.18 GB) beside the Q4_0 GGUF and Ollama's gemma4:12b takes image input; 7 GB fits a 16 GB laptop. No OCR word-error figure for the 12B was found. source
Gemma 4 26B A4B IT BB
Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author)
Google ships gemma-4-26B-it-mmproj.gguf (1.19 GB) beside the Q4_0 GGUF and Ollama's gemma4:26b takes image input; 14.4 GB fits a 24 GB GPU. source
Gemma 4 31B IT BB
Google LLC (the registrant of gemma.ai; the card names Google DeepMind as author)
Google ships gemma-4-31B-it-mmproj.gguf (1.2 GB) beside the Q4_0 GGUF and Ollama's gemma4:31b takes image input; 17.6 GB fits a 24 GB GPU. 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)
Google ships gemma-4-E4B-it-mmproj.gguf (0.99 GB) and Ollama passes images; acc-45 tested Japanese OCR on the family but gives no error rate, and no Tesseract comparison exists. 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)
OpenBMB ships the vision GGUF beside every quant, the card lists OCRBench and OmniDocBench results and high-resolution image input, and 5 to 6 GB of files fit a 16 GB laptop. 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)
Meta ships a 1.4 GB mmproj beside the GGUF and Ollama tags take images, about 22 GB with the build; OmniDocBench 75.8 is Meta's own run and no account tests scans. source
Ornith-1.5-9B C
Ornith
A 0.9 GB mmproj ships beside the GGUFs and Ollama's tag shows vision, but the card gives no document or OCR results. 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)
Publisher GGUF ships the mmproj (0.45 GB), llama.cpp b6907 and the Ollama instruct tag pass images, OCR in 32 languages; no account measures WER or seconds per page at this size. 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)
Publisher GGUF ships the mmproj (1.1 GB), llama.cpp b6907 and the Ollama instruct tag pass images, OCR in 32 languages; no account measures WER or seconds per page at this size. 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)
Publisher GGUF ships the mmproj (0.45 GB), llama.cpp b6907 and the Ollama instruct tag pass images, OCR in 32 languages; no account measures WER or seconds per page at this size. source

Does the job, with a limit 25

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
Vision config present and the publisher's Q4_K_M repository ships Agents-A1-4B-mmproj.gguf (0.67 GB), so llama-server takes images; no OCR result or image test was found and the card documents no vision use. source
Apertus v1.5 8B C
ETH Zurich and EPFL (the Swiss National AI Institute, named in the Acceptable Use Policy)
Takes images (image-text-to-text, vision tokenizer shipped) but no mmproj or stock runtime; a third-party Q4_K_M GGUF converts it to the 1.0 apertus architecture, which cannot carry its image path. source
Apodex-1.1-mini C
Apodex (the site's copyright line reads 2026 Apodex)
The GGUF repository ships mmproj-apodex1.1-35b-F16.gguf and the card's llama-server line passes --mmproj, so images work; 22.6 GB of files need a 24 GB GPU with little room, and no OCR result exists. source
BigBang-v1 C
The Endless Frontier (endless-frontier)
config.json keeps the base's vision tower and bartowski ships an mmproj file, but the card says nothing about images and the publisher ships no GGUF. source
Cosmos-Reason2-2B C
NVIDIA Corporation
Reads images through Qwen3-VL, and a third-party Q4_K_M GGUF ships an mmproj; but NVIDIA states 24 GB minimum GPU memory, ships no quantised build, and trains for physical scenes, not OCR. 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)
Google ships gemma-4-E2B-it-mmproj.gguf (0.99 GB) and Ollama's e2b takes images, but no OCR result for E2B was found and acc-45 ran Japanese OCR on the family without reporting E2B as passing. source
InternVL3.5-GPT-OSS-20B-A4B-Preview D
Shanghai AI Laboratory (OpenGVLab), not named in the repository
Reads images and video through InternViT and vLLM loads it; a third-party MXFP4 GGUF (12 GB) with mmproj fits 16 GB under llama.cpp's gpt-oss text architecture, but it is a stale preview trained at 32K with a template that may be wrong, and no OCR result exists. source
MiMo-V2.6 Distill Qwen 9B C
Xiaomi (the card and report name Xiaomi MiMo; no LICENSE file names a copyright holder)
The Qwen3.5-9B base keeps a vision tower and Ollama's tag carries a vision badge, but Xiaomi ships no GGUF or mmproj and the card claims nothing about documents. source
Nemotron 3 Nano Omni 30B A3B Reasoning C
NVIDIA Corporation
OCRBenchV2 (EN) 65.8 and a PDF-to-image script on the card, Unsloth GGUFs of 18.5 GB to 24 GB with a 1.6 GB mmproj, Ollama tags take images; scores are NVIDIA's and vLLM needs trust_remote_code. source
occamy-1.0 C
Accio (Accio-Lab)
Keeps the base vision tower, a 0.9 GB F16 mmproj ships beside the GGUFs and BF16 passed an image check in vLLM; vision was not retested on the newer quants and no OCR result exists. source
Ornith-1.5-35B-A3B C
Ornith
A 0.9 GB mmproj ships beside the GGUFs and Ollama's tag has a vision badge, but a discussion titled Removed Vision Capabilities is open and the card claims nothing about documents. source
Qwen3.5-0.8B D
Alibaba Cloud (named in the licence's copyright line)
Has a vision encoder, but Alibaba ships no GGUF; the mmproj comes from Unsloth's build, and no account reads documents with this size. source
Qwen3.5-2B D
Alibaba Cloud (named in the licence's copyright line)
Has a vision encoder, but Alibaba ships no GGUF; the mmproj comes from Unsloth's build, and no account reads documents with this size. source
Qwen3.5-4B C
Alibaba Cloud (named in the licence's copyright line)
Vision encoder in config and Ollama's vision badge, but Alibaba ships no GGUF; the mmproj is Unsloth's, and no account measures OCR at this size. source
Qwen3.5-9B C
Alibaba Cloud (named in the licence's copyright line)
Vision encoder and Ollama vision badge, but Alibaba ships no GGUF or mmproj; Unsloth's build carries one, and no account measures OCR on the 9B. source
Qwen3.6-27B B
Alibaba Cloud (named in the licence's copyright line)
Vision encoder, Ollama vision badge and Unsloth mmproj, 17 to 18 GB with projector on a 24 GB GPU; Alibaba ships no GGUF and no OCR account exists. source
Qwen3.6-35B-A3B B
Alibaba Cloud (named in the licence's copyright line)
Vision encoder, Ollama vision badge and Unsloth mmproj, about 23 GB with projector on a 24 GB GPU; Alibaba ships no GGUF and no OCR account exists. source
Qwen3.8-27B B
Alibaba Cloud (named in the licence's copyright line)
Vision encoder, Ollama vision badge and Unsloth mmproj; about 18 GB with projector on a 24 GB GPU, but Alibaba ships no GGUF and no OCR account exists. source
Qwen3-Omni-30B-A3B-Instruct D
Alibaba Cloud (carried from the 20 October 2026 pilot; this model's files name no copyright holder)
Takes images and llama.cpp lists the ggml-org GGUF with its mmproj (18.6 GB Q4 plus 1.3 GB), fitting a 24 GB GPU; Alibaba ships no GGUF and no OCR account exists. 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)
Publisher GGUF ships the mmproj and 19.8 GB Q4 fits a 24 GB GPU at short context; acc-16 OCRed thousands of medical pages locally but does not name the size. 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)
acc-43 OCRed 600 dpi typewriter scans with the 8B on Ollama, better than Tesseract, but at 5.5 minutes per page on a 16 GB CPU laptop; publisher GGUF ships the mmproj. source
Qwopus3.8-27B-Flash-V2 C
Jackrong, an individual Hugging Face account
The GGUF repository ships mmproj-F32 and the base takes images, 17 GB on a 24 GB GPU; no OCR account and a Hermes-style template the card does not describe. source
Swift 1.5 Qwen3.8-27B C
UkisAI
The GGUF ships a vision projector and 17.4 GB Q4 fits a 24 GB GPU, but the MLX 4-bit build has no image chat, the licence has a revenue cap and no OCR account exists. source
ThinkingCap (Qwen3.8 27B) C
BottleCap AI (named in the licence's copyright line)
The GGUF ships an f16 mmproj and the card documents llama-server --mmproj, 17.4 GB fits a 24 GB GPU; licence cap applies and no OCR account exists. source
Baidu Unlimited-OCR-2 C
Baidu, Inc.
Parses scans and rendered PDF pages into text, boxes and HTML tables with 6.7 GB of BF16, but needs trust_remote_code or a repository SGLang wheel, was tested on an A800, and has no publisher GGUF; third-party GGUFs run only in a niche runtime. source

Doesn't do this job 51

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

ListingWhy
Carbon-X C
Hugging Face (the HuggingFaceBio organisation, named Hugging Face Biology Research on the Hub)
Text only; no vision_config. source
Darwin-36B-Opus C
VIDRAFT Inc. (named on vidraft.net); the card's citation names FINAL-Bench and VIDRAFT_LAB
Text only; the base's vision encoder is not in the files. source
DeepSeek-V3.2 C
DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)
Text only; no vision_config. source
DeepSeek V4.1 Flash C
DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)
Takes images (vision_config, mmproj in the Unsloth GGUF) but the smallest build is 233 GB, far past a 16 GB laptop or 24 GB GPU. source
DeepSeek-V4-Flash-0731 C
DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)
Text only; no vision_config. source
DiffusionGemma 26B A4B IT D
Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author)
Takes images, but no released local runtime loads it and Google asks for more than 60 GB of GPU memory. 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)
Text only. 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)
Text only. 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
Text only. source
GLM-5.3-Flash B
Z.AI Co., Ltd (named in the licence's copyright line)
Takes images and Unsloth ships an mmproj, but the smallest build is 93 GB, far past a 16 GB laptop or 24 GB GPU. source
gpt-oss-120b BB
OpenAI (named as publisher on the card; the openai-api listing's terms name OpenAI OpCo, LLC)
Text only. 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)
Text only. source
Granite 4.2 3B C
International Business Machines Corporation (per IBM's naming guidance, as read for granite-guardian)
Text only. source
Hy3 B
Tencent (named in the LICENSE file's copyright line)
Text only. source
Inkling B
Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary)
Takes images, but the smallest build is 270 GB. source
Inkling-Small B
Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary)
Takes images and Unsloth ships an mmproj, but 75 GB minimum is far past a 24 GB GPU. source
Kimi K3 B
Moonshot AI (the licence names Moonshot AI; the Kimi Platform terms name Moonshot AI PTE. LTD.)
Takes images but the smallest build is 466 GB. source
Kolibri-1 C
Aleph Alpha GmbH
Text only. 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
No image input.
Laguna M.1 C
Poolside, Inc. (named in Poolside's Terms of Use; the LICENSE.md copyright line reads Poolside)
Text only; config.json has no vision tower.
Laguna S 2.1 C
Poolside, Inc. (named in Poolside's Terms of Use)
Text only.
Laguna XS 2.1 C
Poolside, Inc. (named in Poolside's Terms of Use)
Text only.
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
Only BF16 from inclusionAI, about 55 GB of files served with vLLM on four GPUs; no GGUF or mmproj, so nothing fits a 16 GB laptop or 24 GB GPU. source
LFM2.5-2.6B D
Liquid AI, Inc. (the licensor named in the LFM Open License)
Text only; LFM2.5-VL is a different model.
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
Text only.
MiniCPM5-2B C
OpenBMB (named in the GitHub LICENSE's copyright line)
Text only; a discussion asking about vision is unanswered.
MiniMax-M2.5 C
MiniMax AI (named in the licences; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)
Text only.
MiniMax-M2.7 C
MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)
Text only; MiniMaxM2ForCausalLM has no vision tower. source
MiniMax-M3.1-Flash C
MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)
Takes images and Unsloth ships mmproj files, but no build fits the 16 GB laptop or 24 GB GPU the use case names. source
Mistral Large 4 C
Mistral AI (RCS Paris 952 418 325)
Pixtral-style vision encoder and a 3.4 GB mmproj ship, but no build fits the 16 GB laptop or 24 GB GPU the use case names. source
Nanbeige4.3-33B C
BOSS Zhipin (Kanzhun Limited), unconfirmed
Text only; LlamaForCausalLM with no vision tower. source
NVIDIA Nemotron 3.5 Bolt E4B 4GB C
NVIDIA Corporation
NVIDIA documents text input only and ships no mmproj beside its GGUF, though the BF16 config keeps Gemma 4's vision_config and Ollama's tag lists Image. source
Nemotron 3.5 Lightning 30B A3B B
NVIDIA Corporation
Text only; Ollama lists text input and the config has no vision tower. source
Nemotron 3 Nano 30B A3B B
NVIDIA Corporation
Text only; Ollama lists text input and the config has no vision tower. source
Nemotron 3 Nano 4B C
NVIDIA Corporation
Text only. source
Nemotron 3 Super 120B A12B B
NVIDIA Corporation
Text only. source
Nemotron 3 Ultra 550B A55B B
NVIDIA Corporation
Text only. source
Olmo 3 7B Think D
The Allen Institute for Artificial Intelligence (Ai2), a 501(c)(3) nonprofit, per its Responsible Use Guidelines
Text only; Olmo3ForCausalLM has no vision tower. source
Ornith-1.5-397B C
Ornith
An mmproj ships, but the smallest build is 244 GB. source
Qwen3-30B-A3B-Instruct-2507 C
Alibaba Cloud (named in the licence's copyright line)
Text only; Qwen3MoeForCausalLM has no vision tower. source
Qwen3-4B-Instruct-2507 C
Alibaba Cloud (named in the licence's copyright line)
Text only; no vision tower in config.json. source
Qwen3.5-122B-A10B B
Alibaba Cloud (named in the licence's copyright line)
Vision encoder present, but no build fits a 16 GB laptop or 24 GB GPU. source
Qwen3.5-397B-A17B B
Alibaba Cloud (named in the licence's copyright line)
Vision encoder present but no build fits a 16 GB laptop or 24 GB GPU. source
Qwen3.8-Flash-Next C
Alibaba Cloud (Singapore) Private Limited, the registrant of qwen.ai. The licence names Qwen as copyright holder
Vision encoder present but no build fits a 16 GB laptop or 24 GB GPU. source
Qwen3-Coder-30B-A3B-Instruct C
Alibaba Cloud (named in the licence's copyright line)
Text only; no vision tower in config.json. source
Qwen3-Coder-Next C
Alibaba Cloud (the Qwen team's employer; no LICENSE file names a copyright holder for this model)
Text only; no vision tower in config.json. source
Recursive Flint C
Recursive Superintelligence, Inc. (the Hugging Face organisation's full name)
Text only; no vision tower in config.json. source
Ripple-65B Preview D
Ricursive Intelligence, Inc.
Text only; no vision tower in config.json. source
SmolLM3-3B C
Hugging Face, Inc.
Text only; no vision tower in config.json. source
Spark-X2.6-4B C
XHToken (the licence's copyright line); the Hugging Face organisation's name is SparkLLM
Text only; no vision tower in config.json. source
Swift 1.5 Qwen3.8-Flash-Next C
UkisAI
Vision projector shipped, but no build fits a 16 GB laptop or 24 GB GPU. source

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

What stops agents finishing

  • GGUF shipped without the mmproj file, so images are ignored
  • Ollama tag loads the Thinking edition instead of Instruct
  • five minutes per page on a 16 GB laptop
  • looping hallucination on repeated glyph patterns

Around it

Before: document-extraction.pdf-to-images, file-storage.list. After: document-extraction.table-from-text, spreadsheets.write-rows.

Where this comes from

Weight 8 of 100 in Open-weight models: 4 of 51 accounts; most vision listings carry no user account

The same as data: /api/v1/use-cases/open-weights.document-vision.json, or the MCP tool get_use_case with id open-weights.document-vision. 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.