Up to date
deepseek-ai Apache 2.0 out 9 months ago
3.39B parameters (6 of 64 experts per token), images and text in, text out, 8K tokens of context.
Verified 2026-10-25 Rank #136 in LLMs · score 94.1 method v1 Sources 3 Downloads 30d 531.3k · #162 of 619 tracked
Changed this week: 1 new quantized build on the Hub (11 in all).2026-10-24
Picked by rule from 29 tracked pieces: the lab's own post, the most shared, a field report, the latest.
The 2 picks above, plus these 23. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.
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 39 Spaces.
Top Spaces running it: merterbak/DeepSeek-OCR-Demo ♥ 513 · prithivMLmods/DeepSeek-OCR-2-Unlimited-OCR ♥ 25 · baohuynhbk14/DeepSeek-OCR-2-DEMO ♥ 10 · mcikalmerdeka/vlms-ocr-playground ♥ 3
Weights only, an estimate: leave room for the KV cache. Method
About 2.0 GB at INT4, 4.1 GB at FP8, 8.1 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.
Estimate params × bytes per param × 1.2 · how we estimate
Weights, every dated article, and repo commits on one rail.
Teal is the weights (2026-01-27), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.
57 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Everything else, folded: the full tables, every source.
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.
| Pipeline tag | image-text-to-text |
|---|---|
| Library | transformers |
| Architectures | DeepseekOCR2ForCausalLM |
| Model type | deepseek_vl_v2 |
| Dtype | bfloat16 |
| Params by dtype | BF16: 3,389,119,360 |
| Size band | M |
| MoE | yes |
| Experts | 64 |
| Experts per token | 6 |
| Context length | 8,192 (config.json max_position_embeddings) |
| Languages | multilingual |
| Papers | arXiv:2601.20552, arXiv:2510.18234 |
| License | Apache 2.0 (apache-2.0) |
| Hub snapshot | 2026-10-25 |
vLLM: supported · DeepseekOCR2ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · DeepseekOCR2ForCausalLM · 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.
Weights-only memory for DeepSeek-OCR-2: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 8.1 GB · Single GPU (<24 GB) |
|---|---|
| FP8 (1 byte/param) | 4.1 GB · Single GPU (<24 GB) |
| INT4 (0.5 bytes/param) | 2.0 GB · Edge (<4 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.
finetune: 36quantized: 11adapter: 9merge: 1
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| unsloth/DeepSeek-OCR-2 | finetune | — | 3.39B | M | apache-2.0 | unsloth | 55 | 556.9k |
| Dogacel/Universal-DeepSeek-OCR-2 | finetune | — | 3.39B | M | apache-2.0 | Dogacel | 12 | 36.6k |
| sabafallah/DeepSeek-OCR-2-GGUF | quantized | GGUF | — | — | — | sabafallah | 13 | 35.2k |
| richarddavison/DeepSeek-OCR-2-FP8 | quantized | -FP8 | 3.39B | M | mit | richarddavison | 5 | 8.1k |
| JamePeng2023/DeepSeek-OCR-2-GGUF | quantized | GGUF | — | — | apache-2.0 | JamePeng2023 | 0 | 5.5k |
| SandLogicTechnologies/DeepSeek-OCR-2-GGUF | quantized | GGUF | — | — | mit | SandLogicTechnologies | 2 | 4.6k |
| forkjoin-ai/deepseek-ocr-2 | finetune | — | 3.39B | M | apache-2.0 | forkjoin-ai | 0 | 4k |
| bartsolutions/DeepSeek-OCR-2-GPTQ-INT8 | quantized | GPTQ | 3.39B | M | mit | bartsolutions | 1 | 3.7k |
| 42ailab/DeepSeek-OCR-2-GGUF | quantized | GGUF | — | — | apache-2.0 | 42ailab | 1 | 1.6k |
| harishforaiandml/DeepSeek-OCR-2-Q4_K_M-GGUF | quantized | GGUF | — | — | apache-2.0 | harishforaiandml | 1 | 723 |
| strangervisionhf/deepseek-ocr-2-transformers-v4.57.1 | finetune | — | 3.39B | M | apache-2.0 | strangervisionhf | 2 | 446 |
| AyoubChLin/deepseek_ocr2_arabic_jsonify | adapter | — | 3.39B | M | apache-2.0 | AyoubChLin | 0 | 422 |
| ajayrathi100699/unsloth_finetune_deepseek | finetune | — | 3.39B | M | apache-2.0 | ajayrathi100699 | 0 | 307 |
| YUNGHUI2024/deepseek-ocr2-chart-v1 | adapter | — | — | — | apache-2.0 | YUNGHUI2024 | 0 | 259 |
| Tcyber/deepseek_ocr_ethiopian_legal_FULL | finetune | — | 3.39B | M | apache-2.0 | Tcyber | 0 | 183 |
| Nightwingkills/deepseek_12k_finetune_full | finetune | — | 3.39B | M | apache-2.0 | Nightwingkills | 0 | 177 |
| Nightwingkills/deepseek-ocr2-finetuned_V3 | finetune | — | 3.39B | M | apache-2.0 | Nightwingkills | 0 | 173 |
| ajayrathi100699/unsloth_finetune_deeping | finetune | — | 3.39B | M | apache-2.0 | ajayrathi100699 | 0 | 170 |
| Nightwingkills/unsloth_finetune_sahi_hai | finetune | — | 3.39B | M | apache-2.0 | Nightwingkills | 0 | 161 |
| hmqid03/deepseek-ocr2-vi | finetune | — | 3.39B | M | apache-2.0 | hmqid03 | 0 | 149 |
| Nightwingkills/deepseek-ocr2-finetuned | finetune | — | 3.39B | M | apache-2.0 | Nightwingkills | 0 | 130 |
| nalatisa/finetuned-ocrd2-merged | finetune | — | 3.39B | M | apache-2.0 | nalatisa | 0 | 122 |
| AutomatosX/AX-DeepSeek-OCR-2-MLX-AXQ-MXFP8 | quantized | MLX | 3.39B | M | apache-2.0 | AutomatosX | 0 | 122 |
| melsiddieg/deepseekocr2-ar | finetune | — | 3.39B | M | apache-2.0 | melsiddieg | 0 | 121 |
| raporto/unsloth_finetune | finetune | — | 3.39B | M | apache-2.0 | raporto | 0 | 116 |
| rifathridoy/bangla_deepseek_ocr_2 | finetune | — | 3.39B | M | apache-2.0 | rifathridoy | 0 | 109 |
| Ovalko/Deepseek-OCR-QnA | finetune | — | 3.39B | M | apache-2.0 | Ovalko | 0 | 107 |
| Felipe2231/canario-amarelo-prod-final | merge | — | — | — | artistic-2.0 | Felipe2231 | 2 | 0 |
| jianchen266/DSO | finetune | — | 3.39B | M | apache-2.0 | jianchen266 | 0 | 88 |
| AutomatosX/AX-DeepSeek-OCR-2-CUDA-AXQ-NVFP4-W4A16 | quantized | NVFP4 | 3.39B | M | apache-2.0 | AutomatosX | 0 | 80 |
| tnkchaseme/DeepSeek-OCR-2-CROHME-LORA | adapter | — | — | — | — | tnkchaseme | 0 | 76 |
| pacogarcia/unsloth_deepseek_finetune | finetune | — | 3.39B | M | apache-2.0 | pacogarcia | 0 | 75 |
| loay404/unsloth_finetune | finetune | — | 3.39B | M | apache-2.0 | loay404 | 0 | 70 |
| TensorVizion/RagSeek-v1 | finetune | — | 3.39B | M | apache-2.0 | TensorVizion | 0 | 70 |
| Inoriqq/deepseek-ocr2-thai-lora-clean | adapter | — | — | — | — | Inoriqq | 0 | 69 |
| kamillkate/unsloth_finetune_TW2 | finetune | — | 3.39B | M | apache-2.0 | kamillkate | 0 | 68 |
| AutomatosX/AX-DeepSeek-OCR-2-MLX-AXQ-6bit | quantized | MLX | 3.39B | M | apache-2.0 | AutomatosX | 0 | 65 |
| mason1998/deepseek-finetune-ocr-kurdish | finetune | — | 3.39B | M | apache-2.0 | mason1998 | 0 | 63 |
| AutomatosX/AX-DeepSeek-OCR-2-CUDA-AXQ-NVFP4-W4A4 | quantized | NVFP4 | 3.39B | M | apache-2.0 | AutomatosX | 0 | 63 |
| nickoo004/deepseek-ocr2-karakalpak-lora | adapter | — | — | — | apache-2.0 | nickoo004 | 0 | 61 |
| Date | Commit | SHA |
|---|---|---|
| 2026-02-03 | Update README.md | aaa02f3 |
| 2026-01-30 | Update README.md | c11b6e9 |
| 2026-01-27 | Update README.md | fdee390 |
| 2026-01-27 | Update README.md | f5a5d18 |
| 2026-01-27 | Update README.md | 4d8e200 |
| 2026-01-27 | Delete .ipynb_checkpoints | 18bdb3c |
| 2026-01-27 | initial commit | db8cee3 |
| 2026-01-27 | Update README.md | b5ac633 |
| 2026-01-27 | Delete .ipynb_checkpoints | 16dca68 |
| 2026-01-27 | initial commit | e6322a2 |
| 2026-01-27 | initial commit | e5f86a3 |
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.
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).
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 15:02 UTC · Last verified 2026-10-25
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.
