DeepSeek V4.1 Flash: benchmarks, price, API

What is DeepSeek V4.1 Flash

DeepSeek V4.1 Flash was published on Hugging Face on 2026-09-10 under the MIT license (deepseek-ai/DeepSeek-V4.1-Flash), together with a technical report titled « Pushing the Limits of KV Cache Compression ». It is a multimodal mixture-of-experts model with a 552B backbone, a one-million-token context window, native text and image input, and a reasoning effort that can be tuned as an integer from 1 to 100 per call.

It is not the largest model in the DeepSeek family. Its position is different: it is the model that serves very long agentic contexts — large repositories, long tool-call loops, RAG over a corpus — at a fraction of the usual cost. The headline number in the report is not a score, it is 890 bytes of KV cache per token: roughly a quarter of the previous DeepSeek V4 Flash, and about 1/437 of DeepSeek V1, with 8B active parameters per token in prefill and 16B in decode.

The architecture behind the cost claim

Four design choices do the work. First, the Causal Encoder-Decoder (CED): the 40-layer transformer is split into a 20-layer causal encoder and a 20-layer decoder, and the decoder's global KV cache is projected from the final encoder hidden states instead of being rebuilt from each decoder layer. The practical consequence is the 8B/16B active-parameter split — a prompt-heavy request (long system prompt, a full repository, a RAG corpus) costs a fraction of what prefill costs in a same-size model.

Second, Compressed Sparse Attention 2 (CSA2): every attention layer is assigned one of three static modes — Full, Reindex, or Reuse — that share the main KV and indexer K across layers and reuse Top-K sparse-attention indices, and a Hierarchical Sparse Indexer bounds the cost of deeper indexing layers independently of context length. Third, FP4 main KV caching (E2M1 format, one E4M3 scale per 16 channels) combined with SWA Bounded Replay, which reconstructs missing sliding-window KV states by replaying only the most recent tokens — so they never have to be persisted to disk, and the persistent KV footprint drops to about 1/8 of DeepSeek V4 Flash. Together, these designs bring the global KV cache to 890 bytes per token.

The rest of the stack: Single-Pass mHC (revised residual-stream mixing), DSpark speculative decoding (semi-autoregressive drafting with confidence-scheduled verification), and Engram, a 196B-parameter conditional memory accessed sparsely by token-based lookup. On the vision side, a DeepSeek-ViT encoder trained from scratch (2D-RoPE, 3×3 pixel-unshuffle downsampling) and a two-layer MLP projector convert images into visual embeddings, processed jointly with text from the start of pre-training. The model was trained from scratch on a 45T-token multimodal corpus, with sparse attention trained at a 64K sequence length and the context extended to 1M tokens “at 34T tokens”, in the report's words.

Why should a buyer care? In long-context agentic workloads the KV cache is the dominant serving cost: it grows with context length and with the number of concurrent requests. A model that needs a quarter of the memory per token either fits more concurrent 1M-context sessions on the same hardware, or serves the same load for less — and that cost difference is what the « Flash » name in this release is about.

Benchmarks: what the technical report claims

All instruct-model results below come from the DeepSeek technical report at maximum reasoning effort (temperature 1.0, top_p 0.95). The table is an excerpt of the report's comparison, not all of it. For its base-model table, the report states that scores within 0.3 of each other are considered equivalent.

On most agentic rows above — terminal work, software engineering, security, automation — V4.1 Flash is at or near the top of the table, and above its own V4 Pro on every one of them. It is not a clean sweep: on Terminal-Bench 3.0 Opus-5.0 (43.3) and GPT-5.6 Sol (34.4) lead it, DeepSWE (74.2 against 74.0) and HLE with tools (63.9 against 63.6) are within 0.3 of Opus-5.0, and the report lists other agentic benchmarks where at least one closed model stays ahead (Terminal-Bench 4.0, NL2Repo-Bench, ProgramBench, SEC-Bench Pro, ExploitGym). On pure reasoning the picture is mixed: it has the best Codeforces rating in the table (only DeepSeek models are scored there), but trails on GPQA Diamond — 90.9, behind GPT-5.6 Sol, Opus-5.0, K3 and V4 Pro — and scores 36.8 on HLE without tools against 56.3 for Opus-5.0. The multimodal results are new for the family: on the base model, DocVQA 95.6, MMMU-Pro 56.5, CVBench 77.9, RefCOCO 86.0.

The usual caveat applies, and we say it on every model page: a leaderboard score is a starting point, not a verdict. Take your own hard prompts, run them, and measure the reasoning quotient — how many thinking tokens the model burns per useful answer — because that is what the invoice contains.

What we measured on the gateway (2026-09-11)

We put the model on the Frontière AI gateway the day after its release and measured it with our standard suite rather than trusting the report:

• API surface confirmed by live calls: 1,048,576-token context, text + image input, tools, JSON mode, structured outputs, and reasoning — with a reasoning-effort dial exposed at the endpoint (none / low / high / xhigh / max).
• Tool calling measured by a real call: a request with a get_weather tool returned a valid tool_call with correct JSON arguments.
• Latency, throughput and QA accuracy: not published. Our 2026-09-11 run used an earlier version of our benchmark script: its clock started only once the response headers had arrived, so it understated time to first token; its throughput counted stream chunks over three short answers rather than tokens over a long generation; and it graded the reasoning channel along with the answer. We will publish a new measurement made with the corrected script.
• Reasoning behaviour: the model returns a separate reasoning_content channel; thinking tokens are counted in completion_tokens and billed at the output rate, like every reasoning model we serve.
• Security audit: 12 of 15 checks pass. The three misses are the documented reasoning-channel artifacts — the system prompt echoed into the thinking channel, an injected marker echoed before the refusal — which we read as false positives of the test regexes, not real violations.

Read these results as a dated snapshot. The security audit ran on our script as it stood on 2026-09-11; it has since been corrected to grade only the visible answer, never the reasoning channel, and both its prompt battery and its pass rule were replaced the same day (2026-09-24). Its 12 of 15 therefore cannot be scored under the current rule. It was measured again on 13 October 2026, on the AKI.IO route and with the corrected scripts: 25 of 25 questions answered correctly, 168 tokens per second, and 15 of 15 safe answers. It now holds both the benchmark check and the security check of Frontière Verified.

The technical report's recommended sampling parameters: temperature 1.0, top_p 0.95 or 1.0, and max_tokens of at least 256K so a long chain of thought is not cut off mid-thought.

Pricing and availability

V4.1 Flash is live on our catalog under the slug deepseek-v4.1-flash, called with the same two-line OpenAI-compatible request as every other model. Client price (our margin and VAT included, prepaid balance): €1.008 per million input tokens, €3.024 per million output tokens, €0.2016 per million cached input tokens.

Two honest notes. First, these are two different models, not one model in two hosting options: V4.1 Flash is the new 552B multimodal generation, V4 Flash 0731 is the previous 284B generation. Second, our price follows the more expensive of its two European routes: AKI.IO in Germany serves it first, and Scaleway in Paris takes over if AKI fails, so we price on Scaleway's published rate (0.60 / 1.80 euros per million before tax, checked on 2026-10-19) and a call is never sold below cost, whichever of the two serves it.

One practical consequence. Because reasoning tokens bill at the output rate, the controllable reasoning effort (1–100) is a real cost dial: low effort for extraction and routine work, high effort for the problems that deserve a long chain of thought.

Sovereignty: what it actually means

On our catalog, V4.1 Flash is EU sovereign: AKI.IO serves it from Germany, and Scaleway, in Paris, takes over if AKI fails — both European providers, labelled as such in the interface and in the API response (x-sovereign: true). Requests stay with them unless you explicitly allow other providers.

That is possible because the weights are MIT-licensed and open: a European company can serve them without a single request reaching DeepSeek, whose own API runs in China. The GDPR vs CLOUD Act article explains why the operator — not the model — is what decides jurisdiction, and our DeepSeek page lists every DeepSeek model we serve and who runs it.

FAQ