Up to date
deepseek-ai MIT out 21 months ago
684B parameters (8 of 256 experts per token), text in, text out, 160K tokens of context. 1 provider serves it from $0.70 per million input tokens.
Verified 2026-10-25 Rank #18 in LLMs · score 119.6 method v1 Sources 4 Downloads 30d 1.5M · #86 of 619 tracked
Picked by rule from 35 tracked pieces: the lab's own post, the most shared, a field report, the latest.
A PR for llama.cpp delivers x2 speed for WASM by optimizing SIMD instructions for qX_K_q8_K and qX_0_q8_0 dot product functions.
Zyphra released ZAYA1-8B, a model that matches DeepSeek-R1 on math benchmarks while using less than 1 billion active parameters.
The 4 picks above, plus these 21. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.
Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.
| # | Cheapest live row | In /1M | Out /1M | Context | Quant | Seen |
|---|---|---|---|---|---|---|
| 1 | Novita | $0.70 | $2.50 | 63K | — | Oct 25 |
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.
Served by 1 API provider; heavily run locally (70 quantized builds); a big base for builders (454 fine-tunes, adapters and merges).
Top Spaces running it: akhaliq/anycoder ♥ 3,321 · pliny-the-prompter/obliteratus ♥ 621 · llamameta/DeepSeek-R1-Chat-Assistant-Web-Search ♥ 260 · hadadrjt/ai ♥ 86
Weights only, an estimate: leave room for the KV cache. Method
About 411 GB at INT4, 821 GB at FP8, 1.64 TB 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
Third-party measurements only; author-reported numbers stay in Reference, marked as such.
56.9. Each tick is a model on this site with a measured Aider Polyglot row; author-reported numbers stay in Reference. Percent (pass_rate_2). edit_format=diff
Source · leaderboard · 2025-01-20
Weights, every dated article, and repo commits on one rail.
Teal is the weights (2025-01-20), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.
524 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.
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| measured Aider Polyglot · Write code with agents | 56.9 percent (pass_rate_2) | edit_format=diff | leaderboard | 2025-01-20 |
| author-reported AIME 2024 · Reason through hard problems | 79.8 percent | DeepSeek R1 | model author | 2026-10-07 |
| author-reported AIME 2024 · Reason through hard problems | 28.9 percent | DeepSeek-R1-Distill-Qwen-1.5B | model author | 2026-10-07 |
| author-reported AIME 2024 · Reason through hard problems | 55.5 percent | DeepSeek-R1-Distill-Qwen-7B | model author | 2026-10-07 |
| author-reported AIME 2024 · Reason through hard problems | 69.7 percent | DeepSeek-R1-Distill-Qwen-14B | model author | 2026-10-07 |
| author-reported AIME 2024 · Reason through hard problems | 72.6 percent | DeepSeek-R1-Distill-Qwen-32B | model author | 2026-10-07 |
| author-reported AIME 2024 · Reason through hard problems | 50.4 percent | DeepSeek-R1-Distill-Llama-8B | model author | 2026-10-07 |
| author-reported AIME 2024 · Reason through hard problems | 70 percent | DeepSeek-R1-Distill-Llama-70B | model author | 2026-10-07 |
| author-reported AIME 2024 · Reason through hard problems | 52.7 | DeepSeek-R1-Distill-Qwen-1.5B | model author | 2026-10-07 |
| author-reported AIME 2024 · Reason through hard problems | 83.3 | DeepSeek-R1-Distill-Qwen-7B | model author | 2026-10-07 |
| author-reported AIME 2024 · Reason through hard problems | 80 | DeepSeek-R1-Distill-Qwen-14B | model author | 2026-10-07 |
| author-reported AIME 2024 · Reason through hard problems | 83.3 | DeepSeek-R1-Distill-Qwen-32B | model author | 2026-10-07 |
| author-reported AIME 2024 · Reason through hard problems | 80 | DeepSeek-R1-Distill-Llama-8B | model author | 2026-10-07 |
| author-reported Aider Polyglot · Write code with agents | 53.3 percent | DeepSeek R1 | model author | 2026-10-07 |
| author-reported AlpacaEval · Chat and follow instructions | 87.6 percent | DeepSeek R1 | model author | 2026-10-07 |
| author-reported Arena-Hard · Chat and follow instructions | 92.3 | DeepSeek R1 | model author | 2026-10-07 |
| author-reported C-Eval · Work across languages | 91.8 | DeepSeek R1 | model author | 2026-10-07 |
| author-reported CLUEWSC (EM) | 92.8 | DeepSeek R1 | model author | 2026-10-07 |
| author-reported CNMO 2024 (Pass@1) | 78.8 percent | DeepSeek R1 | model author | 2026-10-07 |
| author-reported Codeforces · Write code with agents | 96.3 percent | DeepSeek R1 | model author | 2026-10-07 |
| author-reported Codeforces · Write code with agents | 2029 rating | DeepSeek R1 | model author | 2026-10-07 |
| author-reported DROP (3-shot F1) | 92.2 | DeepSeek R1 | model author | 2026-10-07 |
| author-reported FRAMES (Acc.) | 82.5 percent | DeepSeek R1 | model author | 2026-10-07 |
| author-reported GPQA Diamond · Reason through hard problems | 71.5 percent | DeepSeek R1 | model author | 2026-10-07 |
| author-reported GPQA Diamond · Reason through hard problems | 33.8 percent | DeepSeek-R1-Distill-Qwen-1.5B | model author | 2026-10-07 |
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| author-reported GPQA Diamond · Reason through hard problems | 49.1 percent | DeepSeek-R1-Distill-Qwen-7B | model author | 2026-10-07 |
| author-reported GPQA Diamond · Reason through hard problems | 59.1 percent | DeepSeek-R1-Distill-Qwen-14B | model author | 2026-10-07 |
| author-reported IFEval · Chat and follow instructions | 83.3 | DeepSeek R1 | model author | 2026-10-07 |
| author-reported LiveCodeBench · Write code with agents | 65.9 percent | DeepSeek R1 | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 97.3 percent | DeepSeek R1 | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 83.9 percent | DeepSeek-R1-Distill-Qwen-1.5B | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 92.8 percent | DeepSeek-R1-Distill-Qwen-7B | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 93.9 percent | DeepSeek-R1-Distill-Qwen-14B | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 94.3 percent | DeepSeek-R1-Distill-Qwen-32B | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 89.1 percent | DeepSeek-R1-Distill-Llama-8B | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 90.8 percent | DeepSeek R1 | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 92.9 | DeepSeek R1 | model author | 2026-10-07 |
| author-reported MMLU-Pro · Reason through hard problems | 84 | DeepSeek R1 | model author | 2026-10-07 |
| author-reported SWE-bench Verified · Write code with agents | 49.2 | DeepSeek R1 | model author | 2026-10-07 |
| author-reported SimpleQA · Reason through hard problems | 30.1 | DeepSeek R1 | model author | 2026-10-07 |
| author-reported SimpleQA · Reason through hard problems | 63.7 | DeepSeek R1 | model author | 2026-10-07 |
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 | text-generation |
|---|---|
| Library | transformers |
| Architectures | DeepseekV3ForCausalLM |
| Model type | deepseek_v3 |
| Dtype | bfloat16 |
| Params by dtype | BF16: 3,918,786,560 · F32: 15,104 · F8_E4M3: 680,571,043,840 |
| Size band | XXL |
| MoE | yes |
| Experts | 256 |
| Experts per token | 8 |
| Context length | 163,840 (config.json max_position_embeddings) |
| Papers | arXiv:2501.12948 |
| License | MIT (mit) |
| Hub snapshot | 2026-10-25 |
vLLM: supported · DeepseekV3ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · DeepseekV3ForCausalLM · 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-R1: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 1.64 TB · Cluster (≥640 GB) |
|---|---|
| FP8 (1 byte/param) | 821 GB · Cluster (≥640 GB) |
| INT4 (0.5 bytes/param) | 411 GB · Multi-GPU node (<640 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: 70finetune: 327merge: 5adapter: 122
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| unsloth/DeepSeek-R1-GGUF | quantized | GGUF | — | — | mit | unsloth | 1.1k | 8.5M |
| amd/DeepSeek-R1-MXFP4 | quantized | MXFP4 | 371B | XXL | mit | amd | 5 | 1.2M |
| mlx-community/DeepSeek-R1-4bit | quantized | -4bit | 671B | XXL | — | mlx-community | 39 | 400.2k |
| nvidia/DeepSeek-R1-NVFP4 | finetune | NVFP4 | 397B | XXL | mit | nvidia | 286 | 383.2k |
| bartowski/DeepSeek-R1-GGUF | quantized | GGUF | — | — | mit | bartowski | 108 | 229.9k |
| QuixiAI/DeepSeek-R1-AWQ | quantized | AWQ | 671B | XXL | mit | QuixiAI | 87 | 137k |
| mlx-community/DeepSeek-R1-3bit | quantized | — | 671B | XXL | — | mlx-community | 15 | 136k |
| unsloth/DeepSeek-R1-BF16 | finetune | -BF16 | 684B | XXL | mit | unsloth | 24 | 50.9k |
| lmstudio-community/DeepSeek-R1-GGUF | quantized | GGUF | — | — | — | lmstudio-community | 16 | 50.6k |
| microsoft/MAI-DS-R1 | finetune | — | 671B | XXL | mit | microsoft | 314 | 21k |
| tngtech/DeepSeek-TNG-R1T2-Chimera | merge | — | 684B | XXL | mit | tngtech | 287 | 21k |
| silence09/DeepSeek-R1-Small-2layers | finetune | — | 2.43B | M | mit | silence09 | 9 | 34.7k |
| RedHatAI/DeepSeek-R1-quantized.w4a16 | finetune | -quant | 676B | XXL | mit | RedHatAI | 7 | 34.5k |
| unsloth/DeepSeek-R1-GGUF-UD | quantized | GGUF | — | — | mit | unsloth | 16 | 32.4k |
| amd/DeepSeek-R1-MXFP4-ASQ | quantized | MXFP4 | 363B | XXL | mit | amd | 1 | 30k |
| nvidia/DeepSeek-R1-NVFP4-v2 | finetune | NVFP4 | 394B | XXL | mit | nvidia | 10 | 29.1k |
| tngtech/DeepSeek-R1T-Chimera | merge | — | 684B | XXL | mit | tngtech | 272 | 14.7k |
| unsloth/DeepSeek-R1 | finetune | — | 684B | XXL | mit | unsloth | 53 | 20.6k |
| calcuis/deepseek-r1 | quantized | GGUF | — | — | mit | calcuis | 3 | 21.5k |
| Barrrrry/DeepSeek-R1-W4AFP8 | quantized | — | 357B | XXL | mit | Barrrrry | 3 | 21.1k |
| microsoft/MAI-DS-R1-FP8 | quantized | -FP8 | 671B | XXL | mit | microsoft | 31 | 19.7k |
| opensourcerelease/DeepSeek-R1-bf16 | finetune | -bf16 | 684B | XXL | mit | opensourcerelease | 17 | 13.1k |
| XythicK/DeepSeek-R1-GGUF | quantized | GGUF | — | — | — | XythicK | 2 | 11.6k |
| mlx-community/DeepSeek-R1 | quantized | — | — | — | mit | mlx-community | 3 | 6.4k |
| OPEA/DeepSeek-R1-int4-sym-gguf-q4-0-inc | quantized | int4 | — | — | mit | OPEA | 14 | 5.9k |
| persadian/CropSeek-LLM | finetune | — | — | — | other | persadian | 3 | 5.1k |
| rohithsiddhartha/DeepSeek-R1-4bit | quantized | -4bit | — | — | — | rohithsiddhartha | 1 | 4.5k |
| Alphatao/Affine-0000000 | quantized | — | 685B | XXL | mit | Alphatao | 0 | 4.4k |
| mlx-community/DeepSeek-R1-2bit | quantized | — | 671B | XXL | mit | mlx-community | 9 | 3.3k |
| Zarkdrd/Lane | quantized | GGUF | — | — | — | Zarkdrd | 1 | 3.4k |
| RecurvAI/Recurv-Medical-Deepseek-R1 | finetune | — | — | — | mit | RecurvAI | 26 | 1.5k |
| xdatasi/xdata-finetune-deepseek-reason-test-medical | quantized | GGUF | 4.78B | M | apache-2.0 | xdatasi | 1 | 2.6k |
| raghu1155/DeepSeek-R1-Codegeneration-COT | finetune | — | — | — | apache-2.0 | raghu1155 | 11 | 1.9k |
| Open4bits/DeepSeek-R1-mlx-2Bit | quantized | mlx | 671B | XXL | mit | Open4bits | 2 | 2.1k |
| SKT-NRS/SKT-SURYA-H | finetune | — | — | — | apache-2.0 | SKT-NRS | 24 | 847 |
| OPEA/DeepSeek-R1-int4-gptq-sym-inc | quantized | int4 | 671B | XXL | — | OPEA | 6 | 1.7k |
| beita6969/DeepSeek-R1-Distill-Qwen-32B-Medical | finetune | — | 32.8B | L | — | beita6969 | 6 | 1.3k |
| rim89987/DeepSeek-R1-0528-Qwen3-8B-abliterated-Q5_0-GGUF | quantized | Q5_0 | — | — | mit | rim89987 | 1 | 1.4k |
| OPEA/DeepSeek-R1-int2-mixed-sym-inc | quantized | — | 59.4B | XL | — | OPEA | 12 | 676 |
| imsanjoykb/deepSQL-R1-distill-8B | adapter | — | — | — | apache-2.0 | imsanjoykb | 4 | 887 |
| Date | Commit | SHA |
|---|---|---|
| 2025-03-27 | Small fix | 56d4cbb |
| 2025-02-24 | Update bib info | a157fa3 |
| 2025-02-09 | Update tokenizer_config.json | 8a58a13 |
| 2025-02-01 | Update README.md | f7361cd |
| 2025-01-26 | Update README.md | 5dde110 |
| 2025-01-23 | Update README.md | 4dc77d4 |
| 2025-01-22 | Update tokenizer_config.json | cb48aa8 |
| 2025-01-22 | Update tokenizer_config.json | 7bae4ba |
| 2025-01-21 | Update README.md (#16) | 3302ba7 |
| 2025-01-21 | Tag Model as MIT license (#12) | 17e58a0 |
| 2025-01-21 | Add generation_config.json | ebc477a |
| 2025-01-20 | Release DeepSeek-R1 | 5a56bdb |
| 2025-01-20 | Add files using upload-large-folder tool | 952e8ef |
| 2025-01-20 | Add files using upload-large-folder tool | ca135fb |
| 2025-01-20 | Add files using upload-large-folder tool | 1d646a4 |
| 2025-01-20 | Add files using upload-large-folder tool | 2aec6f9 |
| 2025-01-20 | Add files using upload-large-folder tool | 0c37a84 |
| 2025-01-20 | Add files using upload-large-folder tool | 455029a |
| 2025-01-20 | Add files using upload-large-folder tool | 21b6a6f |
| 2025-01-20 | Add files using upload-large-folder tool | abf9bba |
| 2025-01-20 | initial commit | e175219 |
Hosted endpoints observed for DeepSeek-R1, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.70/M.
| Provider | Model ref | Context | Quant | $/M in | $/M out | tok/s | TTFT | Tools | Status | Observed |
|---|---|---|---|---|---|---|---|---|---|---|
| Novita hf | deepseek/deepseek-r1-turbo | 63K | — | $0.70 | $2.50 | 23.3 | 1005 ms | yes | live | 2026-10-25 |
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 21:31 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 14:59 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.
