Up to date
deepseek-ai Not declared out 22 months ago
684B parameters (8 of 256 experts per token), text in, text out, 160K tokens of context. 2 providers serve it from $0.26 per million input tokens.
Verified 2026-10-25 Rank #49 in LLMs · score 108.1 method v1 Sources 4 Downloads 30d 1.4M · #88 of 619 tracked
Picked by rule from 24 tracked pieces: the lab's own post, the most shared, a field report, the latest.
DeepSeek-V3 is a strong Mixture-of-Experts language model with 671B total parameters and 37B activated for each token.
DeepSeek AI released the DeepSeek-V3 model on December 26th, 2024, which beats the pair of GPT-4o and Claude 3.5 together on evaluations considered extremely hard.
A series of performance analyses models DeepSeek-V3 training efficiency from theoretical roofline limits to realistic PyTorch profiles.
The 4 picks above, plus these 20. 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.
2 providers with a live price, from $0.26 to $0.26 per million input tokens; the median is $0.26. One dot per provider, cheapest in teal.
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 2 API providers; 17 quantized builds for local use; a big base for builders (124 fine-tunes, adapters and merges).
Top Spaces running it: akhaliq/anycoder ♥ 3,321 · pliny-the-prompter/obliteratus ♥ 621 · gokaygokay/Chroma ♥ 252 · acvlab/abot-world-interactive ♥ 97
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.
36.67. Each tick is a model on this site with a measured SWE-bench row; author-reported numbers stay in Reference. Percent resolved. agent: KGCompass
Source · leaderboard · 2025-06-09
Weights, every dated article, and repo commits on one rail.
Teal is the weights (2024-12-25), amber triangles are 8 dated articles (each links to it), grey ticks are repo commits.
141 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 LiveCodeBench · Write code with agents | 49.55 percent pass@1 (mean over questions) | all 1055 questions in file, 2023-05-08..2025-04-07 | leaderboard | 2025-04-07 |
| measured SWE-bench · Write code with agents | 36.67 percent resolved | agent: KGCompass | leaderboard | 2025-06-09 |
| measured SWE-bench · Write code with agents | 30.67 percent resolved | agent: Moatless Tools | leaderboard | 2025-01-11 |
| author-reported AGIEval (Acc.) | 79.6 percent | Base Model | model author | 2026-10-07 |
| author-reported ARC Challenge · Chat and follow instructions | 95.3 percent | Base Model | model author | 2026-10-07 |
| author-reported BBH (EM) | 87.5 | Base Model | model author | 2026-10-07 |
| author-reported C-Eval · Work across languages | 90.1 percent | Base Model | model author | 2026-10-07 |
| author-reported C3 (Acc.) | 78.6 percent | Base Model | model author | 2026-10-07 |
| author-reported CCPM (Acc.) | 92 percent | Base Model | model author | 2026-10-07 |
| author-reported CLUEWSC (EM) | 82.7 | Base Model | model author | 2026-10-07 |
| author-reported CMMLU · Work across languages | 88.8 percent | Base Model | model author | 2026-10-07 |
| author-reported CMRC (EM) | 76.3 | Base Model | model author | 2026-10-07 |
| author-reported CRUXEval-I (Acc.) | 67.3 percent | Base Model | model author | 2026-10-07 |
| author-reported CRUXEval-O (Acc.) | 69.8 percent | Base Model | model author | 2026-10-07 |
| author-reported Commonsense QA sets · Chat and follow instructions | 98.9 percent | Base Model | model author | 2026-10-07 |
| author-reported Commonsense QA sets · Chat and follow instructions | 84.7 percent | Base Model | model author | 2026-10-07 |
| author-reported Commonsense QA sets · Chat and follow instructions | 82.9 | Base Model | model author | 2026-10-07 |
| author-reported Commonsense QA sets · Chat and follow instructions | 40 | Base Model | model author | 2026-10-07 |
| author-reported DROP (3-shot F1) | 91.6 | Chat Model | model author | 2026-10-07 |
| author-reported DROP (F1) | 89 | Base Model | model author | 2026-10-07 |
| author-reported GSM8K · Reason through hard problems | 89.3 | Base Model | model author | 2026-10-07 |
| author-reported HellaSwag · Chat and follow instructions | 88.9 percent | Base Model | model author | 2026-10-07 |
| author-reported HumanEval · Write code with agents | 65.2 pass@1 | Base Model | model author | 2026-10-07 |
| author-reported IFEval · Chat and follow instructions | 86.1 | Chat Model | model author | 2026-10-07 |
| author-reported LiveCodeBench · Write code with agents | 19.4 pass@1 | Base Model | model author | 2026-10-07 |
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| author-reported MATH-500 · Reason through hard problems | 61.6 | Base Model | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 90.7 | Base Model | model author | 2026-10-07 |
| author-reported MBPP · Write code with agents | 75.4 pass@1 | Base Model | model author | 2026-10-07 |
| author-reported MGSM · Work across languages | 79.8 | Base Model | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 87.1 percent | Base Model | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 86.2 percent | Base Model | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 88.5 | Chat Model | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 89.1 | Chat Model | model author | 2026-10-07 |
| author-reported MMLU-Pro · Reason through hard problems | 64.4 percent | Base Model | model author | 2026-10-07 |
| author-reported MMLU-Pro · Reason through hard problems | 75.9 | Chat Model | model author | 2026-10-07 |
| author-reported MMMLU · Work across languages | 79.4 percent | Base Model | model author | 2026-10-07 |
| author-reported Pile-test (BPB) | 0.55 other | Base Model | model author | 2026-10-07 |
| author-reported RACE-High (Acc.) | 51.3 percent | Base Model | model author | 2026-10-07 |
| author-reported RACE-Middle (Acc.) | 67.1 percent | Base Model | model author | 2026-10-07 |
| author-reported Winogrande · Chat and follow instructions | 84.9 percent | Base Model | 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:2412.19437 |
| License | Not declared |
| 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-V3: 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: 17finetune: 87adapter: 36merge: 1
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| unsloth/DeepSeek-V3-GGUF | quantized | GGUF | — | — | mit | unsloth | 138 | 308.4k |
| mlx-community/DeepSeek-V3-4bit | quantized | -4bit | 671B | XXL | — | mlx-community | 10 | 305.8k |
| opensourcerelease/DeepSeek-V3-bf16 | finetune | -bf16 | 684B | XXL | — | opensourcerelease | 32 | 263.5k |
| hyper-accel/deepseekv3-export-test | finetune | — | 1.05B | S | mit | hyper-accel | 0 | 167.4k |
| unsloth/DeepSeek-V3 | quantized | — | 685B | XXL | mit | unsloth | 13 | 145.3k |
| bullerwins/DeepSeek-V3-GGUF | quantized | GGUF | — | — | other | bullerwins | 103 | 124.9k |
| QuixiAI/DeepSeek-V3-AWQ | quantized | AWQ | 671B | XXL | mit | QuixiAI | 36 | 31.6k |
| unsloth/DeepSeek-V3-bf16 | finetune | -bf16 | 684B | XXL | mit | unsloth | 16 | 18.3k |
| OPEA/DeepSeek-V3-int4-sym-gptq-inc | quantized | int4 | 671B | XXL | — | OPEA | 18 | 10.2k |
| mmnga/DeepSeek-V3-bf16-gguf | quantized | -bf16 | — | — | other | mmnga | 0 | 9.2k |
| mlx-community/DeepSeek-V3-3bit | quantized | — | 671B | XXL | — | mlx-community | 3 | 7.9k |
| rohithsiddhartha/DeepSeek-V3-4bit | quantized | -4bit | — | — | — | rohithsiddhartha | 0 | 6.3k |
| tflsxyy/DeepSeek-V3-bf16-4layers | finetune | -bf16 | 15.1B | L | — | tflsxyy | 0 | 6k |
| OPEA/DeepSeek-V3-int4-sym-gguf-q4-0-inc | quantized | int4 | — | — | — | OPEA | 3 | 4.6k |
| mlx-community/DeepSeek-V3-3bit-bf16 | quantized | -bf16 | 671B | XXL | — | mlx-community | 2 | 3.4k |
| mmnga/DeepSeek-V3-slice-jp64 | finetune | — | 181B | XXL | other | mmnga | 11 | 2.8k |
| v2ray/DeepSeek-V3-1B-Test | finetune | — | 1.05B | S | mit | v2ray | 3 | 2.6k |
| 0xSero/DeepSeek-V3.2-345B-W3A16 | quantized | — | 1.91B | S | apache-2.0 | 0xSero | 10 | 1.6k |
| inarikami/DeepSeek-V3-int4-TensorRT | finetune | int4 | — | — | — | inarikami | 15 | 801 |
| Sica-Cold-Storage-1/DeepSeek-V3-Abliterated | finetune | — | 671B | XXL | mit | Sica-Cold-Storage-1 | 2 | 956 |
| OPEA/DeepSeek-V3-int4-sym-awq-inc | quantized | int4 | 671B | XXL | — | OPEA | 4 | 847 |
| v2ray/DeepSeek-V3-1B-Test-AWQ | quantized | AWQ | 1.05B | S | mit | v2ray | 2 | 879 |
| huihui-ai/DeepSeek-V3-bf16 | finetune | -bf16 | — | — | apache-2.0 | huihui-ai | 2 | 813 |
| OpenC/HEFT-Qwen | finetune | — | 1.78B | S | mit | OpenC | 8 | 365 |
| mradermacher/DeepSeek-V3-GGUF | finetune | GGUF | — | — | — | mradermacher | 15 | 0 |
| renziify/1.PsychologyTest | adapter | — | — | — | apache-2.0 | renziify | 13 | 0 |
| tflsxyy/DeepSeek-V3-4bit-4layers | quantized | -4bit | 9.47B | L | — | tflsxyy | 1 | 331 |
| mradermacher/DeepSeek-V3-i1-GGUF | finetune | GGUF | — | — | — | mradermacher | 6 | 0 |
| franb23/tarot | finetune | — | 109M | S | mit | franb23 | 1 | 223 |
| Ojttt/deepseekv3_export_test | finetune | — | 1.05B | S | mit | Ojttt | 0 | 178 |
| ckoh04/deepseek-v3-pruned-8experts-mtp | quantized | — | 40.0B | XL | — | ckoh04 | 0 | 122 |
| Trilogix1/DeepSeek-V3-Abliterated | finetune | — | 671B | XXL | other | Trilogix1 | 0 | 115 |
| tflsxyy/DeepSeek-V3-bf16 | finetune | -bf16 | 684B | XXL | — | tflsxyy | 0 | 113 |
| ckoh04/deepseek-v3-pruned-8experts-clean | quantized | — | 37.5B | XL | — | ckoh04 | 0 | 113 |
| ameykaran/DilLeiX | finetune | — | 81.7M | S | — | ameykaran | 0 | 92 |
| yuanjingjing/DeepSeek-V3-Abliterated | finetune | — | 671B | XXL | mit | yuanjingjing | 0 | 67 |
| systemeraswer/DeepSeek-V3-Abliterated | finetune | — | 671B | XXL | mit | systemeraswer | 0 | 58 |
| copedek/DeepSeek-V3-Abliterated | finetune | — | 671B | XXL | mit | copedek | 0 | 55 |
| lhca521/DeepSeek-V3-Abliterated | finetune | — | 671B | XXL | mit | lhca521 | 0 | 53 |
| At-Tawheed/quantum-aurora | adapter | — | — | — | mit | At-Tawheed | 1 | 0 |
| Date | Commit | SHA |
|---|---|---|
| 2025-03-27 | Small fix | e815299 |
| 2025-02-24 | Update bib info | 8651896 |
| 2025-01-24 | add `transformers` library name (enable inference) (#66) | 1d044fd |
| 2025-01-24 | Update README_WEIGHTS.md | 0cf1748 |
| 2025-01-24 | fix the calculation of act params | 108e1e0 |
| 2024-12-30 | Update README.md | 4c1f24c |
| 2024-12-27 | Fix fp8_cast_bf16.py: https://github.com/deepseek-ai/DeepSeek-V3/commit/8f1c9488b53068992f9525fab03b1868e6f7c8c1 | 9672b38 |
| 2024-12-27 | Update README.md with vLLM Support (#8) | 9ffbf9c |
| 2024-12-26 | Release DeepSeek-V3 | dd31960 |
| 2024-12-26 | Add files using upload-large-folder tool | 52af7fb |
| 2024-12-26 | Add files using upload-large-folder tool | 0d9fddb |
| 2024-12-26 | Add files using upload-large-folder tool | 8a89bc2 |
| 2024-12-26 | Add files using upload-large-folder tool | e24a9e9 |
| 2024-12-26 | Add files using upload-large-folder tool | bd0dd75 |
| 2024-12-26 | Add files using upload-large-folder tool | b734e5d |
| 2024-12-26 | Add files using upload-large-folder tool | a5aa653 |
| 2024-12-26 | Add files using upload-large-folder tool | 4b992ad |
| 2024-12-26 | Add files using upload-large-folder tool | d990f04 |
| 2024-12-25 | initial commit | 198edc6 |
Hosted endpoints observed for DeepSeek-V3, 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: $0.26/M.
| Provider | Model ref | Context | Quant | $/M in | $/M out | tok/s | TTFT | Tools | Status | Observed |
|---|---|---|---|---|---|---|---|---|---|---|
| DeepInfra hf, deepinfra, openrouter | deepseek-ai/DeepSeek-V3 | 160K | fp4 | $0.26 | $0.71 | 61.1 | 588 ms | yes | live | 2026-10-25 |
| StreamLake openrouter | deepseek/deepseek-chat | 125K | unknown | $0.26 | $1.03 | — | — | yes | live | 2026-10-24 |
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 21:32 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.
