Up to date
deepseek-ai MIT out 3 months ago
304B parameters (6 of 256 experts per token), text in, text out, 1M tokens of context. 24 providers serve it from $0.042 per million input tokens.
Verified 2026-10-25 Rank #12 in LLMs · score 129.1 method v1 Sources 5 Downloads 30d 4.3M · #34 of 619 tracked
Changed this week: 1 provider no longer listed: Together AI.2026-10-24 · Cheapest input price rose 193.1%: $0.014 → $0.042 per 1M (OpenRouter).2026-10-24 · 1 new quantized build on the Hub (202 in all).2026-10-24
Empty reply.
______
/ \
/ \
\ /
\______/
| |
| |
/| |\
| |
/ \
~~~~~ ~~~~~ ~~~~~
~~~~~~~~~~~~~~~~~~~~No playable frames in the reply.
The ringer: the same four playful tests for every model, replies shown as received and checked by code. A joke, not a benchmark; never ranked.2 of 4 pass · parasail-deepseek-v4-flash-0731 on parasail · run 2026-10-09 · temperature 0 · cost $0.010 · 2026-10 thing: supermarket · all models →
Picked by rule from 45 tracked pieces: the lab's own post, the most shared, a field report, the latest.
DeepSeek V4 Flash 0731 (Max) is a 284B parameter open-weight model released on July 31, 2026, under the MIT license.
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.
22 providers with a live price, from $0.042 to $0.44 per million input tokens; the median is $0.14. 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.
Widely served (24 API providers); heavily run locally (202 quantized builds).
Top Spaces running it: MiniMaxAI/MiniMax-Music3 ♥ 368 · embedl/hfviewer ♥ 181 · victor/DeepSeek-V4-Flash-0731-free-endpoint ♥ 88 · Svngoku/jupyter2agent ♥ 8
Weights only, an estimate: leave room for the KV cache. Method
About 183 GB at INT4, 365 GB at FP8, 730 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
The probe replies as received on 2026-10-07 from parasail, graded by code. Receipts, not a score.
def is_balanced(s: str) -> bool:
stack = []
pairs = {')': '(', ']': '[', '}': '{'}
for ch in s:
if ch in '([{':
stack.append(ch)
elif ch in ')]}':
if not stack or stack.pop() != pairs[ch]:
return False
return not stack amber-falcon-7731
{"city": "Lisbon", "unit": "c"} {"name": "Foo-9B-Instruct", "license": "Apache-2.0", "parameters_b": 9.2, "context_tokens": 131072, "languages": ["English", "French", "German", "Japanese"]} {
"providers": [
{
"provider": "parasail",
"endpoint_model": "parasail-deepseek-v4-flash-0731",
"source_url": "https://api.parasail.io/v1/chat/completions",
"ttft_ms_median": 2918,
"ttft_any_ms_median": 1803,
"tokens_per_s_median": 81,
"runs": [
{
"ttft_s": 2.644,
"ttft_any_s": 1.805,
"tokens_per_s": 36.4,
"latency_s": 8.78,
"tokens_out": 319,
"reasoning_chars": 583,
"finish": "stop"
},
{
"ttft_s": 7.497,
"ttft_any_s": 0.796,
"tokens_per_s": 84.4,
"latency_s": 9.13,
"tokens_out": 770,
"reasoning_chars": 2868,
"finish": "stop"
},
{
"ttft_s": 2.918,
"ttft_any_s": 1.803,
"tokens_per_s": 81,
"latency_s": 3.9,
"tokens_out": 316,
"reasoning_chars": 607,
"finish": "stop"
}
]
}
],
"note": "Single client, wall clock; tok/s counts every output token including hidden reasoning."
} {
"calls": [
{
"name": "get_weather",
"args": {
"city": "Lisbon",
"date": "2026-10-14"
}
},
{
"name": "check_calendar",
"args": {
"date": "2026-10-14"
}
},
{
"name": "book_table",
"args": {
"restaurant": "Cervejaria Ramiro",
"date": "2026-10-14",
"time": "19:30",
"party_size": 2
}
}
],
"checks": {
"get_weather": true,
"check_calendar": true,
"book_table": true
},
"final": "The weather in Lisbon on 2026-10-14 is clear with no rain (5% chance), so I checked your calendar. The free evening slot was 19:30–21:30, and I booked a table for 2 at Cervejaria Ramiro at 19:30.\n\n**Confirmation code: TBL-4471**",
"fixture": "receipts.errand.v1"
} Weights, every dated article, and repo commits on one rail.
Teal is the weights (2026-07-31), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.
248 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 | 25.2 | — | model author | 2026-10-07 |
| author-reported AutomationBench Public | 25.1 | — | model author | 2026-10-07 |
| author-reported Cybergym | 76.7 | — | model author | 2026-10-07 |
| author-reported DSBench-FullStack † | 68.7 | — | model author | 2026-10-07 |
| author-reported DSBench-Hard † | 59.6 | — | model author | 2026-10-07 |
| author-reported DeepSWE · Write code with agents | 54.4 | — | model author | 2026-10-07 |
| author-reported NL2Repo · Write code with agents | 54.2 | — | model author | 2026-10-07 |
| author-reported Terminal-Bench · Write code with agents | 82.7 | — | model author | 2026-10-07 |
| author-reported Toolathlon · Call tools and APIs | 70.3 | — | 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: 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) |
| 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-0731: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 730 GB · Cluster (≥640 GB) |
|---|---|
| FP8 (1 byte/param) | 365 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.
A fixed probe set run against a live API endpoint: the same questions for every model, answers graded by code, nothing judged by a model. Latencies are single-request wall clock from one client location, not a benchmark.
| Probe | Result | Latency | Reply (truncated) |
|---|---|---|---|
| Self-description | info | 1.7 s | I am DeepSeek-V3, developed by DeepSeek (深度求索). My training data cutoff is July 2024. |
| Tool call | pass | 0.8 s | {"city": "Lisbon", "unit": "c"} |
| Strict JSON | pass | 2.5 s | {"name": "Foo-9B-Instruct", "license": "Apache-2.0", "parameters_b": 9.2, "context_tokens": 131072, "languages": ["English", "French", "German", "Japanese"]} |
| Code (hidden tests) | pass | 1.4 s | def is_balanced(s: str) -> bool: stack = [] pairs = {')': '(', ']': '[', '}': '{'} for ch in s: if ch in '([{': stack.append(ch) elif ch in ')]}': if not stack or stack.pop() != pairs[ch]: return False return not stack |
| Multi-step arithmetic | pass | 0.7 s | 16 |
| Needle recall at 128k tokens | pass | 9.6 s | amber-falcon-7731 |
Probe set firstlook.v1 · observed 2026-10-07 20:08 UTC · endpoint · provider docs · cost of this run $0.02
quantized: 202finetune: 38adapter: 7merge: 1
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| huihui-ai/Huihui-DeepSeek-V4-Flash-0731-abliterated-GGUF | quantized | GGUF | — | — | mit | huihui-ai | 227 | 1.3M |
| unsloth/DeepSeek-V4-Flash-0731-GGUF | quantized | GGUF | — | — | mit | unsloth | 780 | 738.4k |
| MJPansa/DeepSeek-V4-Flash-0731-NVFP4 | quantized | NVFP4 | 304B | XXL | mit | MJPansa | 18 | 461.6k |
| AtlasCloud/DeepSeek-V4-Flash-0731-FP8-DSpark | quantized | -FP8 | 304B | XXL | mit | AtlasCloud | 5 | 449.1k |
| 0xSero/deepseek-v4-flash-0731-spark | quantized | — | 60.3B | XL | mit | 0xSero | 51 | 386.3k |
| lmstudio-community/DeepSeek-V4-Flash-0731-GGUF | quantized | GGUF | — | — | mit | lmstudio-community | 7 | 101k |
| drowzeys/keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-Anchored-Tensors | quantized | — | 304B | XXL | other | drowzeys | 129 | 71.3k |
| bartowski/DeepSeek-V4-Flash-0731-GGUF | quantized | GGUF | — | — | mit | bartowski | 41 | 66.9k |
| nvidia/DeepSeek-V4-Flash-0731-NVFP4 | quantized | NVFP4 | 304B | XXL | mit | nvidia | 55 | 49k |
| ggml-org/DeepSeek-V4-Flash-0731-GGUF | quantized | GGUF | — | — | mit | ggml-org | 25 | 38.4k |
| apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8 | finetune | -FP8 | 304B | XXL | mit | apetersson | 38 | 37.5k |
| otheru/DeepSeek-V4-Flash-Strix-Halo-GGUF | quantized | GGUF | — | — | other | otheru | 27 | 32.8k |
| nazeshinjite/DeepSeek-V4-Flash-0731-ds4-GGUF | quantized | GGUF | — | — | mit | nazeshinjite | 21 | 32.9k |
| qtum/DeepSeek-V4-Flash-0731-GGUF | quantized | GGUF | — | — | mit | qtum | 2 | 29.2k |
| Lucebox/DeepSeek-V4-Flash-0731-ROCmFP3 | quantized | GGUF | — | — | mit | Lucebox | 27 | 27.9k |
| AtomicChat/DeepSeek-V4-Flash-0731-GGUF | quantized | GGUF | — | — | mit | AtomicChat | 32 | 25.9k |
| cebeuq/DeepSeek-V4-Flash-0731-abliterated | finetune | — | 306B | XXL | mit | cebeuq | 44 | 20.4k |
| TensorFold/DeepSeek-V4-Flash-0731-MXFP4-MLX | quantized | MXFP4 | 304B | XXL | mit | TensorFold | 23 | 21.5k |
| PrimeIntellect/DeepSeek-V4-Flash-0731-bf16 | finetune | -bf16 | 284B | XXL | mit | PrimeIntellect | 0 | 22.6k |
| Rednalreden/DeepSeek-V4-Flash-0731-dwarfstar-q2-gguf | quantized | gguf | — | — | mit | Rednalreden | 15 | 17.7k |
| lovesenko/DeepSeek-V4-Flash-0731-Abliterated | quantized | — | 304B | XXL | mit | lovesenko | 9 | 17.6k |
| mlx-community/DeepSeek-V4-Flash-0731-2.4bit-mixed | quantized | — | 304B | XXL | mit | mlx-community | 28 | 16.2k |
| dealignai/DeepSeek-V4-Flash-0731-JANG-CRACK | finetune | — | 284B | XXL | mit | dealignai | 33 | 15.4k |
| ox-ox/DeepSeek-V4-Flash-0731-gguf-ds4 | quantized | gguf | — | — | mit | ox-ox | 26 | 14.2k |
| apetersson/DeepSeek-V4-Flash-0731-DS4-Quality128 | quantized | GGUF | — | — | mit | apetersson | 5 | 13.1k |
| unsloth/DeepSeek-V4-Flash-0731 | quantized | — | 304B | XXL | mit | unsloth | 34 | 11k |
| 0xSero/DeepSeek-V4-Flash-Spark-Mini-GGUF | quantized | GGUF | — | — | mit | 0xSero | 9 | 12.1k |
| Jundot/DeepSeek-V4-Flash-0731-oQ4e-mtp | quantized | — | 304B | XXL | mit | Jundot | 8 | 12.1k |
| KeinNiemand/DeepSeek-V4-Flash-0731-IK_GGUF | quantized | GGUF | — | — | mit | KeinNiemand | 8 | 11.8k |
| mlx-community/DeepSeek-V4-Flash-0731-OptiQ-2bit | quantized | — | 304B | XXL | mit | mlx-community | 10 | 11.5k |
| 0xSero/DeepSeek-V4-Flash-0731-REAP | quantized | — | 193B | XXL | mit | 0xSero | 35 | 8.5k |
| bullerwins/DeepSeek-V4-Flash-0731-GGUF | quantized | GGUF | — | — | mit | bullerwins | 9 | 9.7k |
| webbrain-one/DeepSeek-V4-Flash-0731-Vision-BF16 | quantized | -BF16 | 305B | XXL | other | webbrain-one | 6 | 8.8k |
| schlaflos/DeepSeek-V4-Flash-0731-antirez-ds4-GGUF | quantized | GGUF | — | — | mit | schlaflos | 10 | 8k |
| alessandrobologna/DeepSeek-V4-Flash-0731-DSpark-Drafter-GGUF | quantized | GGUF | — | — | mit | alessandrobologna | 3 | 8.3k |
| ddalcu/DeepSeek-V4-Flash-0731-MLX-Serve-mixed-2-3-8bit | quantized | MLX | 324B | XXL | mit | ddalcu | 10 | 7.8k |
| True2456/DeepSeek-V4-Flash-0731-REAP37-native-MLX | quantized | MLX | 30.6B | L | mit | True2456 | 15 | 6.9k |
| squanchyzx/DeepSeek-V4-Flash-0731-HERETIC-Abliterated-FP8 | quantized | -FP8 | 305B | XXL | mit | squanchyzx | 21 | 6.4k |
| 0xSero/DeepSeek-V4-Flash-Spark-GGUF | quantized | GGUF | — | — | mit | 0xSero | 12 | 6.7k |
| chuzka/Huihui-DeepSeek-V4-Flash-0731-abliterated-GGUF | quantized | GGUF | — | — | mit | chuzka | 0 | 6.3k |
Hosted endpoints observed for DeepSeek-V4-Flash-0731, 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: $0.04/M.
| Provider | Model ref | Context | Quant | $/M in | $/M out | tok/s | TTFT | Tools | Status | Observed |
|---|---|---|---|---|---|---|---|---|---|---|
| Relace openrouter | deepseek/deepseek-v4-flash-0731 | 1M | fp4 | $0.04 | $0.18 | — | — | yes | live | 2026-10-24 |
| Baidu openrouter | deepseek/deepseek-v4-flash-0731 | 1M | fp8 | $0.04 | $0.13 | — | — | yes | live | 2026-10-24 |
| Reka openrouter | deepseek/deepseek-v4-flash-0731 | 256K | unknown | $0.05 | $0.53 | — | — | yes | live | 2026-10-24 |
| DeepInfra hf, deepinfra, openrouter | deepseek-ai/DeepSeek-V4-Flash-0731 | 1M | fp8 | $0.06 | $0.18 | 49.7 | 427 ms | yes | live | 2026-10-25 |
| StreamLake openrouter | deepseek/deepseek-v4-flash-0731 | 1000K | fp8 | $0.09 | $0.26 | — | — | yes | live | 2026-10-24 |
| Inceptron openrouter | deepseek/deepseek-v4-flash-0731 | 1M | fp4 | $0.11 | $0.37 | — | — | yes | live | 2026-10-24 |
| DigitalOcean openrouter | deepseek/deepseek-v4-flash-0731 | 1M | unknown | $0.12 | $0.24 | — | — | yes | live | 2026-10-24 |
| Baseten hf, openrouter | deepseek-ai/DeepSeek-V4-Flash-0731 | 1M | fp8 | $0.13 | $0.26 | 163.7 | 227 ms | yes | live | 2026-10-25 |
| Wafer openrouter | deepseek/deepseek-v4-flash-0731 | 1M | nvfp4 | $0.13 | $0.23 | — | — | yes | live | 2026-10-24 |
| Venice openrouter | deepseek/deepseek-v4-flash-0731 | 977K | unknown | $0.13 | $0.26 | — | — | yes | live | 2026-10-24 |
| CoreWeave openrouter | deepseek/deepseek-v4-flash-0731 | 256K | fp8 | $0.13 | $0.28 | — | — | yes | live | 2026-10-24 |
| Parasail parasail, openrouter | parasail-deepseek-v4-flash-0731 | 1M | fp8 | $0.14 | $0.28 | — | — | yes | serverless | 2026-10-24 |
| Cohere openrouter | deepseek/deepseek-v4-flash-0731 | 1M | unknown | $0.14 | $0.28 | — | — | yes | live | 2026-10-24 |
| Sail Research openrouter | deepseek/deepseek-v4-flash-0731 | 1M | fp4 | $0.15 | $0.15 | — | — | yes | live | 2026-10-24 |
| Alibaba openrouter | deepseek/deepseek-v4-flash-0731 | 977K | unknown | $0.18 | $0.53 | — | — | yes | live | 2026-10-24 |
| Mancer 2 openrouter | deepseek/deepseek-v4-flash-0731 | 1M | fp8 | $0.20 | $0.60 | — | — | yes | live | 2026-10-24 |
| SiliconFlow openrouter | deepseek/deepseek-v4-flash-0731 | 1M | fp8 | $0.22 | $0.66 | — | — | yes | live | 2026-10-24 |
| GMICloud openrouter | deepseek/deepseek-v4-flash-0731 | 1024K | fp8 | $0.29 | $0.86 | — | — | yes | live | 2026-10-24 |
| Phala openrouter | deepseek/deepseek-v4-flash-0731 | 1M | unknown | $0.31 | $0.92 | — | — | yes | live | 2026-10-24 |
| Novita hf, openrouter | deepseek/deepseek-v4-flash-0731 | 1M | fp8 | $0.41 | $1.23 | 89.9 | 729 ms | yes | live | 2026-10-25 |
| AtlasCloud openrouter | deepseek/deepseek-v4-flash-0731 | 1M | fp4 | $0.44 | $1.32 | — | — | yes | live | 2026-10-24 |
| Cloudflare openrouter | deepseek/deepseek-v4-flash-0731 | 1M | unknown | $0.44 | $1.32 | — | — | yes | live | 2026-10-24 |
| Featherless hf | deepseek-ai/DeepSeek-V4-Flash-0731 | — | — | — | — | — | — | — | live | 2026-10-25 |
| Scaleway hf | deepseek-v4-flash-0731 | — | — | — | — | 151.4 | 606 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.