Up to date
deepseek-ai MIT out 4 months ago
165B parameters (6 of 256 experts per token), text in, text out, 1M tokens of context.
Verified 2026-10-25 Rank #384 in LLMs · score 79.9 method v1 Sources 3 Downloads 30d 148.6k · #272 of 619 tracked
Picked by rule from 27 tracked pieces: the lab's own post, the most shared, a field report, the latest.
Cactus releases Needle 3, a foundation model for mobile, wearables, robots, smart home, automotive and microcontrollers.
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: 13 quantized builds.
Weights only, an estimate: leave room for the KV cache. Method
About 99 GB at INT4, 198 GB at FP8, 397 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 (2026-06-27), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.
17 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 |
|---|---|---|---|---|
| author-reported AGIEval (EM) | 82.6 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported BBH (EM) | 86.9 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported BigCodeBench · Write code with agents | 56.8 pass@1 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported C-Eval · Work across languages | 92.1 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported CLUEWSC (EM) | 82.2 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported CMMLU · Work across languages | 90.4 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported Commonsense QA sets · Chat and follow instructions | 82.8 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported DROP (F1) | 88.6 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported FACTS Parametric (EM) | 33.9 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported GSM8K · Reason through hard problems | 90.8 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported HellaSwag · Chat and follow instructions | 85.7 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported HumanEval · Write code with agents | 69.5 pass@1 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported LongBench v2 · Read long documents | 44.7 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 57.4 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 93.6 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported MGSM · Work across languages | 85.7 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 88.7 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 89.4 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported MMLU-Pro · Reason through hard problems | 68.3 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported MMLU-Pro · Reason through hard problems | 87.5 | DeepSeek-V4-Pro Max | model author | 2026-10-07 |
| author-reported MMMLU · Work across languages | 88.8 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported MultiLoKo (EM) | 42.2 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported SimpleQA · Reason through hard problems | 30.1 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported SuperGPQA · Reason through hard problems | 46.5 | DeepSeek-V4-Flash-Base | model author | 2026-10-07 |
| author-reported Winogrande · Chat and follow instructions | 79.5 | DeepSeek-V4-Flash-Base | 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 | DeepseekV4ForCausalLM |
| Model type | deepseek_v4 |
| Dtype | bfloat16 |
| Params by dtype | BF16: 1,483,567,488 · F32: 37,741,630 · F8_E4M3: 6,304,038,912 · I8: 148,176,371,712 · I64: 2,327,040 · F8_E8M0: 9,261,408,000 |
| Size band | XXL |
| MoE | yes |
| Experts | 256 |
| Experts per token | 6 |
| Context length | 1,048,576 (config.json max_position_embeddings) |
| Papers | arXiv:2606.19348 |
| License | MIT (mit) |
| Hub snapshot | 2026-10-25 |
vLLM: supported · DeepseekV4ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · DeepseekV4ForCausalLM · 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-V4-Flash-DSpark: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 397 GB · Multi-GPU node (<640 GB) |
|---|---|
| FP8 (1 byte/param) | 198 GB · Multi-GPU node (<640 GB) |
| INT4 (0.5 bytes/param) | 99 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: 13finetune: 4
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 21:39 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:09 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.
