The story so far
Weights, every dated article, and repo commits on one rail.
Most coverage came in the first two weeks
Teal is the weights (2025-11-18), amber triangles are 3 dated articles (each links to it), grey ticks are repo commits.
HunyuanOCR
tencent tencent-hunyuan-community out 11 months ago
1.12B parameters, images and text in, text out, 128K tokens of context.
Verified 2026-10-25 Rank #172 in LLMs · score 90.8 method v1 Sources 3 Downloads 30d 844.3k · #127 of 619 tracked
Can I use it?tencent-hunyuan-communityCustom license; read the repo LICENSE before commercial hostinglicence file2026-10-24
Do people use it?844.3kdownloads, 30 days#127 of 619 we track · top 21%+5 likes, downloads −5% this week841 likesHugging Face2026-10-24
Where it's used
Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.
No API provider serves it, so its downloads are people running it themselves: 11 quantized builds and 19 Spaces.
- As a service0no API provider listed
- On own hardware11quantized builds
- Built on97 fine-tunes · 2 adapters · 0 merges
- In apps19Hugging Face Spaces
Top Spaces running it: tencent/HunyuanOCR ♥ 81 · victor/HunyuanOCR-EN ♥ 7
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits a 16 GB card at INT4
About 0.7 GB at INT4, 1.3 GB at FP8, 2.7 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
What people built on it
20 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
20 derivatives on the Hub, most of them quantized
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Reference
Everything else, folded: the full tables, every source.
What the card says
1 cited sentence
Use-case evidence
6 use cases
What it is good at 6 use cases
Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.
- Understand images and videostructuraldecision-model estimate p=0.99tag: image-text-to-textpipeline: image-text-to-text+1 more
- Extract text from documentsstructuraldecision-model estimate p=1.00name: OCRtag: ocr
- Chat and follow instructionsstructuraltag: conversationalchat template present
- Run on a laptop or phonestructuraldecision-model estimate p=0.97fits: edge
- Read long documentsstructuralconfig: 131,072 context
- Work across languagesstructuraltag: multilingual
Facts
22 fields
Listed in LLMs
Facts
- HunyuanOCR has 1.12B parameters. (safetensors metadata, observed 2026-10-24)
- Context length is 131,072 tokens (128K). (config.json max_position_embeddings, observed 2026-10-24)
- License: tencent-hunyuan-community (friction: review). Custom license; read the repo LICENSE before commercial hosting. (license file, observed 2026-10-24)
- Created 2025-11-18, last modified 2026-08-28. (HF API, observed 2026-10-24)
| Pipeline tag | image-text-to-text |
|---|---|
| Library | transformers |
| Architectures | HunYuanVLForConditionalGeneration |
| Model type | hunyuan_vl |
| Dtype | bfloat16 |
| Params by dtype | BF16: 1,119,925,744 |
| Size band | S |
| MoE | no |
| Context length | 131,072 (config.json max_position_embeddings) |
| Languages | multilingual, en, zh |
| Papers | arXiv:2607.04884 |
| License | tencent-hunyuan-community (other) |
| Hub snapshot | 2026-10-25 |
Servability
vLLM · SGLang
Servability
vLLM: supported · HunYuanVLForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed
vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.
Hardware fit
3 estimates
Hardware fit ESTIMATE
Weights-only memory for HunyuanOCR: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 2.7 GB · Edge (<4 GB) |
|---|---|
| FP8 (1 byte/param) | 1.3 GB · Edge (<4 GB) |
| INT4 (0.5 bytes/param) | 0.7 GB · Edge (<4 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.
All variants
20 variants
Lineage
quantized: 11finetune: 7adapter: 2
Top variants (20 shown, ranked by downloads)
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| ggml-org/HunyuanOCR-GGUF | quantized | GGUF | — | — | — | ggml-org | 15 | 35.7k |
| prithivMLmods/HunyuanOCR-1.5-GGUF-Updated | quantized | GGUF | — | — | apache-2.0 | prithivMLmods | 4 | 13.8k |
| mradermacher/HunyuanOCR-GGUF | quantized | GGUF | — | — | other | mradermacher | 1 | 10.4k |
| mradermacher/HunyuanOCR-i1-GGUF | quantized | GGUF | — | — | other | mradermacher | 1 | 6.7k |
| AnandSingh/hunyuanocr-GGUF | quantized | GGUF | — | — | other | AnandSingh | 0 | 1.6k |
| bombman/HunyuanOCR-GGUF | quantized | GGUF | — | — | apache-2.0 | bombman | 0 | 1.5k |
| AnandSingh/hunyuanocr-mlx | finetune | mlx | 1.12B | S | other | AnandSingh | 2 | 870 |
| SandLogicTechnologies/hunyuanocr-GGUF | quantized | GGUF | — | — | apache-2.0 | SandLogicTechnologies | 0 | 666 |
| hadeseus/HunyuanOCR-mlx-8bit | quantized | mlx | 996M | S | other | hadeseus | 1 | 553 |
| hadeseus/HunyuanOCR-mlx | finetune | mlx | 996M | S | other | hadeseus | 0 | 523 |
| hadeseus/HunyuanOCR-mlx-4bit | quantized | mlx | 996M | S | other | hadeseus | 0 | 436 |
| metanthropic/MahenOCR-1B | finetune | — | 996M | S | apache-2.0 | metanthropic | 0 | 307 |
| jaqja/HunYuanOCR-SFT | finetune | — | 996M | S | — | jaqja | 0 | 210 |
| XxxxYyyy/HunyuanOCR-Q8_0-GGUF | quantized | Q8_0 | — | — | other | XxxxYyyy | 0 | 149 |
| JokerAn/HunyuanOCR | finetune | — | 996M | S | other | JokerAn | 0 | 98 |
| xiaobaicai520/HunyuanOCR | finetune | — | 996M | S | other | xiaobaicai520 | 0 | 73 |
| thiyagesh/sarakku_kadai | finetune | — | — | — | intel-research | thiyagesh | 0 | 0 |
| Yooghrn/Cincau1 | adapter | — | — | — | bigscience-openrail-m | Yooghrn | 0 | 0 |
| shantipriya/hunyuan-ocr-odia | adapter | — | — | — | apache-2.0 | shantipriya | 0 | 0 |
| xxzigou/HunyuanOCR-ncnn | quantized | — | — | — | apache-2.0 | xxzigou | 0 | 0 |
Repository changelog
30 commits
Repository changelog 30 commits
| Date | Commit | SHA |
|---|---|---|
| 2026-08-28 | Update config.json | 47644ec |
| 2026-08-26 | Update config.json | b7bf724 |
| 2026-07-28 | update README | 449e7d4 |
| 2026-07-22 | update README | a8d9d9c |
| 2026-07-16 | Update generation_config.json | de8f10a |
| 2026-07-13 | Update config.json | 3d3cbc9 |
| 2026-07-07 | Update README.md | 9e01f89 |
| 2026-07-07 | Update README.md | d42c2cd |
| 2026-07-06 | README: drop preview notices; scope default OCR prompt to document parsing (#32) | 97160b6 |
| 2026-07-06 | README: 1.0/1.5 layout note; scope default OCR prompt to document parsing (#31) | 6503295 |
| 2026-07-06 | Add HunyuanOCR-1.5 (target at root, DFlash under dflash/, archive 1.0 under v1.0/) (#30) | 01b3ec8 |
| 2026-01-13 | Update README (#26) | f6af82e |
| 2025-12-24 | Update tokenizer_config.json | fed9438 |
| 2025-12-23 | Update tokenizer_config.json | 350e760 |
| 2025-12-17 | Update tokenizer_config.json | 51c1a22 |
| 2025-12-02 | update README.md (#22) | 987eba7 |
| 2025-11-27 | Update README (#15) | 1ab3786 |
| 2025-11-27 | Update README (#14) | 06f6dc1 |
| 2025-11-26 | Update README (#11) | 9a0c818 |
| 2025-11-26 | update README (#8) | e268223 |
| 2025-11-26 | Update README.md (#7) | e412ba1 |
| 2025-11-26 | Update README (#6) | 7502740 |
| 2025-11-26 | Update README (#5) | 7d53690 |
| 2025-11-25 | Update README.md | 885259f |
| 2025-11-25 | Update README.md | d420772 |
| 2025-11-25 | Update generation_config.json | ff93505 |
| 2025-11-25 | Change Title link (#2) | ea397eb |
| 2025-11-25 | Update README (#1) | b28a172 |
| 2025-11-25 | Update README.md | c3bf532 |
| 2025-11-25 | Update README.md | 5bf2a1f |
Articles and coverage
6 pieces
Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Analysis6
arxiv.orgauthority17,384 words
HunyuanOCR-1.5 is a lightweight, end-to-end OCR vision-language model that upgrades its predecessor to be faster and better. Powered by DFlash, it achieves a 6.37x speedup in Transformer inference and a 2.14x speedup under vLLM. The model extends to 4K resolution and a 128K context window while ranking top-tier on OmniDocBench v1.6.written, reviewed, cited
- HunyuanOCR-1.5 supports PC-side inference through llama.cpp, enabling deployment on CPUs, consumer GPUs, and laptops.
- Agentic Data Flow allows agents to deeply participate in material search, tool-based verification, sample cleaning, and data pipeline development.
github.comauthority2,149 words
HunyuanOCR-1.5 is a lightweight, end-to-end OCR-specialized vision-language model that unifies document parsing, text spotting, information extraction, and text-image translation within a single end-to-end VLM. The model achieves better performance via an Agentic Data Flow system and an upgraded training recipe where pretraining Stage-3 incorporates multi-image data and historical OCR data with maximum image resolution extended to 4K and context window extended to 128K. It also supports faster inference through DFlash speculative decoding and PC-side deployment via llama.cpp.written, reviewed, cited
- CHAOS-Bench probes OCR VLMs by injecting character-level corruptions into academic-paper images.
- The unified inference environment requires CUDA 13 and runs vLLM, DFlash, and transformers from one install.
github.comauthority2,149 words
Tencent released HunyuanOCR-1.5, a lightweight end-to-end OCR vision-language model designed to be faster and better than its predecessor. The upgrade introduces DFlash speculative decoding for acceleration and supports deployment on PCs via llama.cpp alongside server-grade vLLM options. New benchmarks include CHAOS-Bench for character-level hallucination, while training extends the context window to 128K and maximum image resolution to 4K.written, reviewed, cited
- The unified inference environment requires CUDA 13 and runs vLLM, DFlash, and transformers from a single install.
- DFlash speculative decoding uses a lightweight block-diffusion draft model to verify candidate tokens in a single pass.
- The Agentic Data Flow system translates model weaknesses into executable data requirements through agent-driven material search and verification.
curateclick.com2025-11-25HN 1 pts · 1 comments1,396 words
HunyuanOCR is a 1B parameter multimodal model from Tencent that unifies detection, recognition, parsing, and translation into a single inference pipeline. It requires 80GB GPU memory for 16K-token decoding and achieves an OmniDocBench overall score of 94.10, outperforming modular and general VLMs on document parsing tasks. vLLM is the recommended deployment path as it delivers better throughput and latency today compared to the Transformers path.written, reviewed, cited
- Readme scripts default to bfloat16 and device_map="auto", requiring deliberate memory sharding in multi-GPU setups.
blog.omagiclee.com2025-11-16
emergentmind.com2025-11-244,247 words
Papers
1 paper
Papers 1
arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).
Similar models
3 models
Similar models
- LightOnOCR-2-1B · 1.01B same size band (S), same task (image-text-to-text), more providers, permissive license
- Qwen3.5-0.8B · 873M same size band (S), same task (image-text-to-text), more providers, permissive license
- MiniCPM-V-4.6 · 1.30B same size band (S), same task (image-text-to-text), permissive license
Sources
3 sources
Sources
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:35 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:03 UTC · Last verified 2026-10-25
Follow HunyuanOCR
Watch this model
- /feeds/models/tencent--HunyuanOCR.xml(RSS 2.0) · every change to HunyuanOCR
- /feeds/families/tencent--HunyuanOCR.xml(RSS 2.0) · every change across the HunyuanOCR family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.