DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash, published on 10 September 2026 under the MIT licence, is an attack on the running cost of long conversations rather than on the leaderboards. Its headline number is not an accuracy score but a memory figure: 890 bytes of key-value cache per token, about a quarter of what its own predecessor needed six weeks earlier, and by DeepSeek's count 437 times less than the company's first model. That matters because the key-value cache is what an assistant must keep in memory to remember what has already been said. In a million-token session it is the dominant cost, and it is the reason long-context models are expensive to serve rather than expensive to train. DeepSeek attacks it from several directions at once: a causal encoder-decoder layout in which twenty decoder layers project their shared memory from the encoder's final states instead of building their own; an attention scheme that gives each layer one of three fixed roles so that layers can share and reuse index structures; and storage of that memory in four-bit floating point. The result is a model that activates 8 billion parameters while reading a prompt and 16 billion while writing an answer, out of 552 billion in the backbone — where the larger V4-Pro activates 49 billion out of 1.6 trillion. The savings reach the price list. Through DeepSeek's API the model costs $0.30 per million input tokens at peak and $1.20 per million output tokens, with off-peak hours charged at exactly half, against $0.44 and $1.32 for the model it replaces. It has also taken over the plain name `deepseek-flash`, with the older identifiers still accepted. On the vendor's own comparison table it leads the field on several agentic tests: 90.6 on Terminal-Bench 2.1, 74.2 resolved on DeepSWE v1.1, 88.1 on CyberGym, 54.8 on AutomationBench and a Codeforces rating of 3,471 — each above Opus-5.0, GPT-5.6 Sol and its own larger sibling. A reader comparing models should look one row further down the same table. On Terminal-Bench 3.0 and 4.0 — the newer, harder revisions of the test it wins at version 2.1 — it scores 30.0 and 31.2 against Opus-5.0's 43.3 and 51.8. On Humanity's Last Exam it reaches 36.8 against 56.3. The pattern is consistent: where a benchmark is mature and the field is crowded at the top, this model is at the front; where the test still has room to discriminate, the gap to the most expensive Western models is wide. It is a superb instrument for cost-sensitive agent work and not a claim to the frontier of reasoning. The model is natively multimodal, reading images through a vision encoder DeepSeek trained from scratch, and was pre-trained on 45 trillion multimodal tokens. Reasoning effort is a dial from 1 to 100 rather than a switch. Weights, a technical report and the licence are published together; adoption in the 30 days to 16 September 2026 stands at 140,636 Hugging Face downloads.

▸News

DeepSeek's new model beats Opus and GPT on the agent test everyone has already saturated. On the two harder versions of that same test, it loses by twenty points

9/12/2026 · DeepSeek model card, DeepSeek-V4.1-Flash

▸Videos

No videos yet.

▸Reviews