Up to date
deepseek-ai MIT out 12 months ago
3.34B parameters (6 of 64 experts per token), images and text in, text out, 8K tokens of context.
Verified 2026-10-25 Rank #55 in LLMs · score 106.5 method v1 Sources 3 Downloads 30d 2.1M · #61 of 619 tracked
Picked by rule from 41 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 100 Spaces.
Top Spaces running it: khang119966/DeepSeek-OCR-DEMO ♥ 456 · embedl/hfviewer ♥ 181 · prithivMLmods/Super-OCRs-Demo ♥ 44 · akhaliq/DeepSeek-OCR ♥ 16
Weights only, an estimate: leave room for the KV cache. Method
About 2.0 GB at INT4, 4.0 GB at FP8, 8.0 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 (2025-10-17), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.
172 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 | DeepseekOCRForCausalLM |
| Model type | deepseek_vl_v2 |
| Dtype | bfloat16 |
| Params by dtype | BF16: 3,336,106,240 |
| Size band | M |
| MoE | yes |
| Experts | 64 |
| Experts per token | 6 |
| Context length | 8,192 (config.json max_position_embeddings) |
| Languages | multilingual |
| Papers | arXiv:2510.18234 |
| License | MIT (mit) |
| Hub snapshot | 2026-10-25 |
vLLM: supported · DeepseekOCRForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · DeepseekOCRForCausalLM · 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: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 8.0 GB · Single GPU (<24 GB) |
|---|---|
| FP8 (1 byte/param) | 4.0 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.
quantized: 11finetune: 131adapter: 27merge: 3
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| NexaAI/DeepSeek-OCR-GGUF | quantized | GGUF | — | — | — | NexaAI | 52 | 211.6k |
| unsloth/DeepSeek-OCR | finetune | — | 3.34B | M | mit | unsloth | 41 | 82.1k |
| prithivMLmods/DeepSeek-OCR-Latest-BF16.I64 | finetune | -BF16 | 3.34B | M | mit | prithivMLmods | 9 | 59.4k |
| SamMikaelson/deepseek-ocr-qvlm-4bit | finetune | -4bit | 1.90B | S | apache-2.0 | SamMikaelson | 0 | 48k |
| sabafallah/DeepSeek-OCR-GGUF | quantized | GGUF | — | — | — | sabafallah | 13 | 26.6k |
| NexaAI/DeepSeek-OCR-GGUF-CUDA | quantized | GGUF | — | — | — | NexaAI | 8 | 21.1k |
| Jalea96/DeepSeek-OCR-bnb-4bit-NF4 | quantized | -bnb | 3.43B | M | mit | Jalea96 | 17 | 20.3k |
| Dogacel/DeepSeek-OCR-Metal-MPS | finetune | — | 3.34B | M | mit | Dogacel | 13 | 2.6k |
| sanchezalonsodavid17/DeepSeek-OCR-MBQ-Quantized-v1 | quantized | -Quant | 3.43B | M | mit | sanchezalonsodavid17 | 6 | 2.4k |
| forkjoin-ai/deepseek-ocr-gguf | quantized | gguf | — | — | apache-2.0 | forkjoin-ai | 0 | 998 |
| larshiakzemil/unsloth_finetune | quantized | GGUF | 8.77B | M | apache-2.0 | larshiakzemil | 0 | 981 |
| neuralabs/deepseek_ocr_de | finetune | — | 3.34B | M | apache-2.0 | neuralabs | 5 | 403 |
| sthaps/DeepSeek-ocr-Thai | adapter | — | — | — | apache-2.0 | sthaps | 2 | 488 |
| prithivMLmods/DeepSeek-OCR-Latest-BF16.I64-v2.0 | finetune | -BF16 | 3.34B | M | apache-2.0 | prithivMLmods | 1 | 512 |
| see-k/unsloth_finetune | quantized | GGUF | 3.34B | M | apache-2.0 | see-k | 0 | 424 |
| arpitingle/sanskrit-ocr-lora | adapter | — | — | — | — | arpitingle | 4 | 157 |
| vinhnx90/deepseek-orc-3b-persian-detection-f16-vllm | finetune | — | 3.34B | M | apache-2.0 | vinhnx90 | 2 | 187 |
| zakir0101/ftv6-ocr_b16 | finetune | — | 3.34B | M | apache-2.0 | zakir0101 | 0 | 278 |
| suyu-io/DeepSeek-OCR-Chinese-3B-Merged | adapter | — | 3.34B | M | apache-2.0 | suyu-io | 1 | 224 |
| melsiddieg/deepseek_ocr_arabic_v2 | finetune | — | 3.34B | M | apache-2.0 | melsiddieg | 1 | 211 |
| akmaldira/DeepSeek-OCR-RKLLM | finetune | — | — | — | mit | akmaldira | 2 | 150 |
| zerostratos/unsloth_finetune_deepseek-ocr | finetune | — | 3.34B | M | apache-2.0 | zerostratos | 0 | 242 |
| amd/DeepSeek-OCR-MXFP4 | quantized | MXFP4 | 2.20B | M | mit | amd | 0 | 241 |
| KrorngAI/deepseek_ocr_Khmer_finetuned | finetune | — | 3.34B | M | apache-2.0 | KrorngAI | 0 | 241 |
| mingyi456/DeepSeek-OCR-DF11 | quantized | — | 4.52B | M | mit | mingyi456 | 1 | 183 |
| not-lain/finetuned_deepseek_ocr | adapter | — | 3.34B | M | mit | not-lain | 1 | 174 |
| SamMikaelson/deepseek-ocr-mbq-w4bit | finetune | — | 3.34B | M | mit | SamMikaelson | 0 | 222 |
| JeffersonNunn/deepseek-ocr-metal | finetune | — | — | — | mit | JeffersonNunn | 4 | 0 |
| CrazyDentisit/DeepSeek-OCR-Chinese-image-unsloth_finetune | finetune | — | 3.34B | M | apache-2.0 | CrazyDentisit | 0 | 197 |
| Yagna1/deepseek-ocr-telugu-chitralekha | finetune | — | 3.34B | M | apache-2.0 | Yagna1 | 0 | 181 |
| melsiddieg/deepseek_ocr_arabic_v6 | finetune | — | 3.34B | M | apache-2.0 | melsiddieg | 1 | 127 |
| Luca207/deepseek_ocr_finetuned2 | finetune | — | 3.34B | M | apache-2.0 | Luca207 | 0 | 168 |
| melsiddieg/deepseek_ocr_arabic | finetune | — | 3.34B | M | apache-2.0 | melsiddieg | 0 | 167 |
| melsiddieg/deepseek_ocr_arabic_v3 | finetune | — | 3.34B | M | apache-2.0 | melsiddieg | 0 | 166 |
| GaborMadarasz/deepseek_ocr_unsloth_finetune | finetune | — | 3.34B | M | apache-2.0 | GaborMadarasz | 0 | 165 |
| arad1367/Deepseek_OCR_FL_1000Samples_3B | finetune | — | — | — | apache-2.0 | arad1367 | 0 | 158 |
| Luke-Bergen/Mineral-Nano-1 | finetune | — | 2.65B | M | apache-2.0 | Luke-Bergen | 3 | 0 |
| matica0902/MLX-Video-OCR-DeepSeek-Apple-Silicon | finetune | MLX | — | — | agpl-3.0 | matica0902 | 3 | 0 |
| vng08/DeepSeek-OCR-word-final | finetune | — | 3.34B | M | apache-2.0 | vng08 | 0 | 150 |
| melsiddieg/deepseek_ocr_arabic_v5 | finetune | — | 3.34B | M | apache-2.0 | melsiddieg | 0 | 149 |
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 21:33 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:00 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.
