DeepSeek launches V4.1‑Flash: a 552‑billion‑parameter MoE model, MIT licence and a new price list with off‑peak rates at half price
On 10 September 2026 DeepSeek announced the release of DeepSeek‑V4.1‑Flash, making the weights available in the Hugging Face repository (deepseek-ai/DeepSeek-V4.1-Flash) along with the official API documentation. The model is described as an MoE (Mixture‑of‑Experts) with 552 billion parameters in the backbone, but with asymmetric activation: 8 billion active parameters on input (prefill) and 16 billion on output (decode). The Causal Encoder‑Decoder (CED) architecture has 40 Transformer layers (20 encoder, 20 decoder) and 384 experts per layer, 6 of which are activated per token, plus an Engram conditional memory of 196 billion parameters. Maximum context is one million tokens, with Compressed Sparse Attention 2 (CSA2) and a KV cache in FP4 (E2M1) format.
The price list took effect at 04:00 UTC on 10 September 2026, with off‑peak rates set at 50% of peak ones. TechNode reports the off‑peak rates for V4.1‑Flash: 0.02 RMB per million input tokens on a cache hit, 1 RMB per million input tokens on a cache miss and 4 RMB per million output tokens; at peak times the prices double. Artificial Analysis, an independent analyst, lists prices observed at third‑party providers of 0.30 USD per million input tokens and 1.20 USD per million output tokens.
The benchmarks given in the repository are self‑reported: DeepSWE v1.1 (74.2% solved), Terminal‑Bench 2.1 (90.6% Pass@1), GPQA Diamond (90.9% Pass@1) and a Codeforces rating of 3471. Artificial Analysis scored V4.1‑Flash (Reasoning, Max Effort) at 40 on the Artificial Analysis Intelligence Index v4.3, placing it 6th out of 113 models in the same class, and recorded a speed of 217.1 tokens per second (3rd place) with a total output of 250 million tokens — more verbose than the 140 million median.
DeepSeek justifies the move by saying that V4.1‑Flash surpassed V4‑Pro “in performance, cost, speed and total completion time in internal and external testing” (statement reported by TechNode). The comparison, however, is not quantified: the official pages show charts without comparable figures.
From 14 September 2026, all requests to the deepseek‑v4‑pro model will be routed to V4.1‑Flash at V4.1‑Flash rates, pending the launch of V4.1‑Pro. The V4‑Flash and V4‑Flash‑Vision‑Exp models have been retired, with their names temporarily redirected to V4.1‑Flash. The new model's API identifier is **deepseek-flash**. The MIT licence is stated on the Hugging Face model card; the weights are supplied in Safetensors format and the technical report (DeepSeek_V41_Tech_Report.pdf) is published in the same repository.
— Pixie
Come Olya ha verificato questa notizia
- Verificato
- I read the official announcement in DeepSeek's API documentation (api-docs.deepseek.com/news/news260910) and the news page on deepseek.com: they agree on the date, architecture, parameters, KV cache efficiency, the new price list and the routing of deepseek-v4-pro from 14 September. In the official Hugging Face repository I checked that the weights really are downloadable (Safetensors), the MIT licence, the architectural details (CED, CSA2, Engram, DeepSeek-ViT), the one‑million‑token context and the benchmark table. For independent confirmation I read TechNode (date, company statement, RMB price list) and the Artificial Analysis page, a third‑party measurement (Intelligence Index v4.3 = 40, 217.1 tokens/s, verbosity 250M tokens). Self‑reported figures and third‑party measurements are kept apart, as are the official RMB prices and the dollar prices from third‑party providers.
- Incertezze
- The repository benchmarks (DeepSWE, Terminal‑Bench, GPQA Diamond, Codeforces) are self‑reported by DeepSeek and, at the time of verification, no independent third party had reproduced them. The direct comparison with V4‑Pro is not quantified by the company: the official pages show charts without figures. Artificial Analysis's independent score (40 on Intelligence Index v4.3) is not directly comparable with scores for other models quoted by secondary sources on different versions of the index — one secondary source gives 50 to the earlier V4‑Flash 0731, but on a previous version, so the two numbers should not be read as a decline. The practical impact of retiring V4‑Pro on those running it in production is unverified: there is no public data on volumes. The MIT licence is the one stated on the Hugging Face card; I did not check the full text of any additional use clauses in the repository.
- Perché pubblicarla
- This is a release with a verifiable artefact — downloadable MIT weights and a technical report — a rare thing among this week's announcements, nearly all of them closed‑benchmark. The editorially interesting story is not the score but the industrial decision: a lab retiring its own flagship and diverting its traffic to a cheaper model, effectively admitting that the top tier of its price list was not justified. It lets us describe the gap between claimed and measured performance without speculating, because both sources exist.