DeepSeek V4.1 Flash: what it is, and where it runs

DeepSeek V4.1 Flash at a glance

DeepSeek released the weights with a technical report titled “Pushing the Limits of KV Cache Compression”. The figures below come from the model card, read on 6 October 2026.

Where the savings come from

In long agentic work — a whole repository, a long chain of tool calls, retrieval over a corpus — the KV cache is what makes serving expensive. That is the memory a model keeps for every token of context, and it grows with the context and with every concurrent session. V4.1 Flash attacks it four ways:

• A causal encoder-decoder. The decoder's global KV cache is projected from the encoder's final states instead of being built layer by layer, which is what lets only 8B parameters work during prefill: prompt-heavy requests get cheaper to read.
• Compressed Sparse Attention 2. Each attention layer gets one of three fixed modes — Full, Reindex or Reuse — so layers share keys and reuse sparse-attention indices, and a hierarchical indexer caps the cost of deeper indexing whatever the context length.
• An FP4 main KV cache (E2M1 values, one E4M3 scale per 16 channels). With the two designs above, it brings the global cache to 890 bytes per token: about four times less than DeepSeek V4 Flash, and 437 times less than DeepSeek V1.
• SWA Bounded Replay. Sliding-window states are rebuilt by replaying only the latest tokens, so they never need writing to SSD, and the persistent cache falls to about an eighth of V4 Flash's.

Around these sit Single-Pass mHC, a revised way of mixing the residual stream, DSpark speculative decoding, and a vision encoder trained from scratch, DeepSeek-ViT, whose image embeddings were learned jointly with text from the start of pre-training. The model saw 45 trillion tokens; sparse attention was trained at a 64K sequence length, and the context was extended to one million tokens at the 34-trillion mark.

What the report claims, and where it trails

Every instruct score in the report uses the maximum reasoning effort, 100, with temperature 1.0 and top_p 0.95. The table below is an excerpt; the column names are the report's own, and every score is DeepSeek's, not a measurement made by Evanesca.

Read the rows, not the bold. V4.1 Flash leads on Terminal-Bench 2.1, CyberGym, AutomationBench and HLE with tools, and edges Opus-5.0 on DeepSWE by 0.2 points. It trails clearly on the harder agentic sets — Terminal-Bench 3.0 and 4.0, NL2Repo-Bench and, outside this excerpt, ProgramBench, SEC-Bench Pro and ExploitGym. On pure reasoning it sits behind four of the report's columns on GPQA Diamond and scores 36.8 on HLE without tools, against 56.3 for Opus-5.0; its Codeforces rating, 3471, is the report's highest, in a row where only DeepSeek's models are rated.

A leaderboard is a starting point. Run your own hard prompts and count the reasoning tokens a model spends per useful answer: on Evanesca, as anywhere, those tokens are billed as output. The benchmarks page explains what Evanesca checks on each model's route.

Where it runs on Evanesca

On Evanesca, DeepSeek V4.1 Flash is run by AKI.IO — AKI.IO GmbH, a company in Berlin. AKI.IO says that “inference requests are currently processed on GPU infrastructure in German data centers”, and that “prompts, inputs, and generated outputs are processed in volatile memory and are not logged, stored, or analyzed by AKI.IO”. Your prompts go neither to DeepSeek nor to China. Elsewhere, V4.1 Flash is often reached through routers such as OpenRouter, a US company that picks a host for each request — its hosts include companies in the United States, Europe and China, DeepSeek itself among them — and the CLOUD Act can reach US companies wherever their servers are.

Its model page shows 6 of its 6 checks passed on this route. DeepSeek V4 Flash, run by Scaleway in Paris, has passed 6. Sovereignty explains why the operator, not the model, decides which law applies.

Price, and the effort dial

A quick question — 300 tokens in, 400 out — costs about €0.0015 on V4.1 Flash and €0.0007 on V4 Flash, before reasoning tokens. Reasoning tokens are billed as output wherever you run a model, so the effort setting the V4.1 Flash model card describes, from 1 to 100, doubles as a cost dial: low for extraction and routine work, high for problems that deserve a long chain of thought. Open-weight reasoning models compares what thinking costs from one model to the next.

DeepSeek recommends temperature 1.0, top_p 0.95 or 1.0, and a max_tokens of at least 256K, so that a long chain of thought is not cut off. On Evanesca there is no subscription: credit from €2 — see pricing.

V4.1 Flash or V4 Flash?

Both are on Evanesca. V4.1 Flash is the newer generation: it reads images and its context is four times longer. Each has its own page, with its checks: DeepSeek V4.1 Flash, run in Germany, and DeepSeek V4 Flash, hosted in France.

The model is not the DeepSeek app

MIT-licensed weights mean any operator may run the model, and the operator decides where your words go. DeepSeek's own app and API are run by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., whose privacy policy, updated on 10 February 2026, states: “To provide you with our services, we directly collect, process and store your Personal Data in People's Republic of China.”

The same weights, run by AKI.IO in Germany, send nothing to DeepSeek — and neither does V4 Flash, run by Scaleway in Paris. That is what the MIT licence allows: any operator may serve the model, and the operator decides which law your words fall under.

Frequently asked questions

Read next

• Tool calling on open-weight models: why it depends on who runs themTool calling turns a model into an agent, but it depends on the serving stack as much as on the weights. How to check it in one call, and which models pass.
• Open-weight reasoning models: what thinking costs, and how to choose oneReasoning models think before they answer, and that thinking is billed as output. Which open-weight models reason, what it costs, and how to choose one.
• Open-weight AI models in 2026: which to use for what, and who runs themThe open-weight models of 2026 — GLM, Qwen, DeepSeek, GPT-OSS, Llama, Mistral — sorted by use, with size, licence, price and the company that runs them.