The most consequential number in DeepSeek V4.1-Flash, released on September 10, 2026, is not its parameter count or its benchmark score. It is 890 bytes per token — the model's global key-value cache footprint. DeepSeek reports that figure represents roughly one quarter of the HBM and one eighth of the SSD storage required by the previous V4-Flash generation. For anyone running agents that hold a million tokens of context across hours of work, that is the line item that decides whether the deployment is viable.
The weights are published on Hugging Face under an MIT license, with a technical report alongside, and commercial use is permitted.
The Architecture
V4.1-Flash is a 552-billion-parameter Mixture-of-Experts model, but the working set is far smaller. It uses a causal encoder-decoder architecture in which only 8B parameters activate on prefill and 16B on decode. Each MoE layer carries 384 routed experts plus one shared expert, with each token routed to just six of the routed experts. DeepSeek also describes a Compressed Expert Dispatch mechanism handling that routing.
The parameter accounting is worth reading carefully, because published figures differ depending on what is being counted: 552B backbone parameters, approximately 196B Engram conditional-memory parameters, and roughly 763B parameters in total. Headline comparisons against dense models are therefore close to meaningless — the relevant comparison is activated parameters per token, where V4.1-Flash is extraordinarily lean.
Capabilities are broad for an open-weight release: native multimodal vision input, a 1M-token context window, up to 384K output tokens, reasoning, function calling, tool use, structured outputs and prompt caching. The API serves it as `deepseek-flash`.
Price and Performance
On the Artificial Analysis Intelligence Index, V4.1-Flash scores 40 in reasoning at maximum effort — well above the median of 18 for open-weight models of comparable size — and generates output at about 214 tokens per second on DeepSeek's own API.
Pricing is where the release becomes genuinely disruptive. Off-peak API rates are $0.15 per million input tokens and $0.60 per million output, with cache reads at $0.003 per million; peak hours run at double those rates. Even at peak, that sits roughly an order of magnitude or more below the current frontier closed models. Reporting places peak pricing around 30x cheaper than GPT-6 Astra and Claude Fable 5.1.
That is not a claim of parity. An Intelligence Index score of 40 is strong for its weight class, not frontier-leading. The proposition is different: for the enormous share of agentic work that is tool-calling, retrieval, summarisation and routine reasoning over long context, V4.1-Flash may be adequate at a fraction of the token cost — and the memory efficiency means it stays adequate as sessions lengthen.
Why It Matters
Agent economics have been dominated by a quadratic problem. Most agentic systems resend accumulated context on every turn, so token consumption grows far faster than task complexity. The industry's answer has largely been software — compaction, summarisation, rolling windows, semantic caching. V4.1-Flash attacks the same problem in the architecture, by making the cache that holds that context dramatically cheaper to keep resident.
Three consequences follow:
- Self-hosting long-context agents becomes plausible for organisations that could not previously justify the HBM. A quarter of the memory for the same context length changes the hardware bill of materials, not just the API invoice.
- The open-weight tier is now competitive on the dimension enterprises care about most — cost per completed task over long horizons, rather than single-shot benchmark scores.
- Data-residency arguments get cheaper to win. An MIT-licensed model that can be run inside a VPC removes the retention and jurisdiction objections that block managed agent services in regulated sectors.
The Caveats for Self-Hosters
Two practical obstacles temper the "just run it yourself" enthusiasm. The model card states the release does not include a Jinja-format chat template, so teams must handle prompt encoding themselves — a small task that nonetheless produces silently degraded output when done wrong. And official quantised formats for consumer hardware were not broadly available at launch, which makes multi-GPU servers or rented cloud GPU instances the realistic near-term path rather than a workstation.
In other words, the memory savings are real but they are savings against a large baseline. This is a model that gets datacentre-class deployment within reach of mid-sized engineering teams; it is not a laptop model.
Where It Lands in the September Wave
The release closes an unusually dense first ten days of the month, in which five frontier launches shipped: Claude Fable 5.1 and its trusted-access twin on September 1, GPT-6 Astra on September 3, Gemini 3.8 Flash with a defenders-only cyber variant on September 2, Meta's Muse Spark 1.3 the same day, and V4.1-Flash on September 10. Release trackers recorded no new frontier models in the seven days ending September 17 — the wave has, for the moment, broken.
What distinguishes DeepSeek's entry is that it competes on a different axis from the rest of that cohort. The American and European labs shipped capability, much of it gated behind trusted-access tiers because of cyber-offence concerns. DeepSeek shipped efficiency, under a permissive licence, with the weights in public. Both are strategies. Only one of them lets a mid-market engineering team read the technical report, download the artefact and run the thing behind its own firewall this quarter.
