DeepSeek Ships V4.1-Flash, a 552B MoE Model Under MIT License

DeepSeek released V4.1-Flash on Sept. 10, a 552-billion-parameter mixture-of-experts model, the company said in a post on its site. The weights are open on Hugging Face under the MIT license, and the model retires both V4-Flash and V4-Flash-Vision-Exp.

The backbone grew 94 percent, but activation dropped

V4.1-Flash's 552-billion-parameter backbone is roughly a 94 percent jump from V4-Flash's 284 billion, according to VentureBeat's review of the release. Despite the larger backbone, VentureBeat's comparison shows prefill activation falling from 13 billion to 8 billion parameters per token, while decode activation rises from 13 billion to 16 billion. The model card on Hugging Face describes the architecture as a causal encoder-decoder: a 20-layer encoder feeding a 20-layer decoder, with 384 routed experts per layer and 6 activated per token. Context runs to 1 million tokens.

The key-value cache drops to 890 bytes a token

DeepSeek said the new architecture cuts the model's key-value cache, the memory a model holds onto to avoid recomputing earlier context, to 890 bytes per token, about a quarter of V4-Flash's footprint, and to roughly an eighth of the prior generation's SSD storage need. That is a vendor figure DeepSeek has not had independently verified.

Off-peak output costs 60 cents per million tokens

DeepSeek's off-peak API pricing is $0.003 per million cached input tokens, $0.15 per million uncached input tokens and $0.60 per million output tokens, doubling to $0.006, $0.30 and $1.20 during weekday peak windows of 01:00 to 04:00 and 06:00 to 10:00 UTC, VentureBeat reported. Peak output pricing for the V4-Pro model V4.1-Flash replaces ran $3.96 per million tokens, SiliconANGLE reported, putting the new peak rate about 70 percent lower.

Independent testers put it mid-pack, not on top everywhere

Artificial Analysis, a third-party benchmarking group, scored the model 40 on its Intelligence Index in maximum-reasoning mode, ranking sixth of 113 comparable open-weight models against a median score of 18. On DeepSeek's own DeepSWE v1.1 coding benchmark, V4.1-Flash scored 74.2 against Claude Opus 5's 74.0 and GPT-5.6 Sol's 73.0, a difference too small to call a clear win. The model trails on other tasks: Claude Opus 5 leads 43.3 to 30.0 on Terminal-Bench 3.0 and 51.8 to 31.2 on Terminal-Bench 4.0, per DeepSeek's own published charts.

Benchmark	V4.1-Flash	Claude Opus 5	GPT-5.6 Sol
DeepSWE v1.1	74.2	74.0	73.0
Terminal-Bench 3.0	30.0	43.3	n/a
Terminal-Bench 4.0	31.2	51.8	n/a

DeepSeek did not say which harness version or run count produced its charted scores, and the company has not published raw transcripts for the Terminal-Bench runs. The model is downloadable now, with no waitlist or gated form, and Hugging Face lists more than 75,000 downloads of the repository in the past month. Its immediate open-weight rival on size, Qwen's newest reasoning line, and the closed GPT-6 Astra that DeepSeek's benchmarks compare against both remain the models V4.1-Flash will be measured against as more independent testing arrives.