Up to date
deepseek-ai MIT out 55 days ago
305B parameters (6 of 256 experts per token), images and text in, text out, 1M tokens of context. 3 providers serve it from $0.22 per million input tokens.
Verified 2026-10-25 Rank #137 in LLMs · score 94 method v1 Sources 4 Downloads 30d 345.8k · #201 of 619 tracked
Changed this week: 1 new quantized build on the Hub (29 in all).2026-10-24
Picked by rule from 33 tracked pieces: the lab's own post, the most shared, a field report, the latest.
The deepseek-flash model accepts images alongside text for tasks like describing pictures or analyzing charts.
Cactus releases Needle 3, a foundation model for mobile, wearables, robots, smart home, automotive and microcontrollers.
The 3 picks above, plus these 22. 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.
3 providers with a live price, from $0.22 to $0.44 per million input tokens; the median is $0.44. 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 3 API providers; 29 quantized builds for local use.
Weights only, an estimate: leave room for the KV cache. Method
About 183 GB at INT4, 366 GB at FP8, 731 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 (2026-08-31), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.
44 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 Agents' Last Exam | 27.3 | Multimodal Agent Capabilities | model author | 2026-10-07 |
| author-reported ApexBench (Pass@1) | 36.5 pass@1 | Multimodal Agent Capabilities | model author | 2026-10-07 |
| author-reported AutomationBench (Public) | 25.7 | Text Agent Capabilities | model author | 2026-10-07 |
| author-reported Chartography | 64.3 | Multimodal Agent Capabilities | model author | 2026-10-07 |
| author-reported Cybergym | 75.3 | Text Agent Capabilities | model author | 2026-10-07 |
| author-reported DSBench-Hard | 63.6 | Text Agent Capabilities | model author | 2026-10-07 |
| author-reported DeepSWE · Write code with agents | 59.3 | Text Agent Capabilities | model author | 2026-10-07 |
| author-reported NL2Repo · Write code with agents | 57.7 | Text Agent Capabilities | model author | 2026-10-07 |
| author-reported Terminal-Bench · Write code with agents | 83.9 | Text Agent Capabilities | model author | 2026-10-07 |
| author-reported Toolathlon · Call tools and APIs | 75.9 | Text Agent Capabilities | model author | 2026-10-07 |
| author-reported ZeroBench (Pass@5) | 35 pass@5 | Multimodal Agent Capabilities | model author | 2026-10-07 |
| author-reported deep_swe (datacurve/deep-swe) | 59.3 | — | Model Card | 2026-10-24 |
| author-reported terminalbench_2_1 (harborframework/terminal-bench-2.1) | 83.9 | — | Model Card | 2026-10-24 |
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 | DeepseekV4ForCausalLM |
| Model type | deepseek_v4 |
| Dtype | bfloat16 |
| Params by dtype | BF16: 1,949,960,576 · F32: 37,754,174 · F8_E4M3: 6,304,038,912 · I8: 296,352,743,424 · I64: 2,327,040 |
| Size band | XXL |
| MoE | yes |
| Experts | 256 |
| Experts per token | 6 |
| Context length | 1,048,576 (config.json max_position_embeddings) |
| 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-Vision-Exp: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 731 GB · Cluster (≥640 GB) |
|---|---|
| FP8 (1 byte/param) | 366 GB · Multi-GPU node (<640 GB) |
| INT4 (0.5 bytes/param) | 183 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: 29finetune: 14merge: 1
| Date | Commit | SHA |
|---|---|---|
| 2026-09-01 | Updata vLLM recipe (#14) | 6821d6a |
| 2026-09-01 | Update Model Card for SGLang support (#12) | 31ea111 |
| 2026-08-31 | Add community evaluation results for DEEP-SWE, TERMINAL-BENCH-2.1 (#5) | e46e16b |
| 2026-08-31 | Update README.md | 26d7203 |
| 2026-08-31 | Add thinking-mode encoding example and change default interactive temperature to 1.0 | 86f746b |
| 2026-08-31 | Update README.md | 0cd0d27 |
| 2026-08-31 | Update README.md | 729b25b |
| 2026-08-31 | init | 47bede8 |
| 2026-08-31 | Add files using upload-large-folder tool | aa1acf9 |
| 2026-08-31 | initial commit | 1dac146 |
Hosted endpoints observed for DeepSeek-V4-Flash-Vision-Exp, 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: $0.22/M.
| Provider | Model ref | Context | Quant | $/M in | $/M out | tok/s | TTFT | Tools | Status | Observed |
|---|---|---|---|---|---|---|---|---|---|---|
| DeepInfra hf, deepinfra, openrouter | deepseek-ai/DeepSeek-V4-Flash-Vision-Exp | 1M | fp8 | $0.22 | $0.65 | 69 | 470 ms | yes | live | 2026-10-25 |
| GMICloud openrouter | deepseek/deepseek-v4-flash-vision-exp | 1024K | fp8 | $0.44 | $1.32 | — | — | yes | live | 2026-10-24 |
| SiliconFlow openrouter | deepseek/deepseek-v4-flash-vision-exp | 1M | fp8 | $0.44 | $1.32 | — | — | 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.
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:34 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:01 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.
