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.

Specification	Value
Architecture	Causal Encoder-Decoder, 40 layers (20 causal encoder + 20 decoder)
Backbone parameters	552B — MoE: 384 routed experts (6 active per token) + 1 shared expert + 196B Engram conditional memory
Active per token	8B in prefill / 16B in decode
Context	1M tokens
Modalities	text + image in, text out
Reasoning effort	continuously controllable, integer 1–100
License	MIT
Released	2026-09-10

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 64K and the context extended to 1M over 34T tokens.

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). DeepSeek's own convention, stated in the report: scores within 0.3 of each other are considered equivalent.

Benchmark	Opus-5.0	GPT-5.6 Sol	K3	GLM-5.3	V4 Pro	V4 Flash	V4.1 Flash
GPQA Diamond (Pass@1)	93.4	94.1	92.9	88.1	92.4	89.9	90.9
Codeforces (rating)	—	—	—	—	3348	3289	3471
Terminal-Bench 2.1 (Pass@1)	89.1	88.8	88.3	88.2	87.9	82.7	90.6
Terminal-Bench 3.0 (Pass@1)	43.3	34.4	17.7	28.3	11.8	7.6	30.0
DeepSWE v1.1 (Resolved)	74.0	73.0	67.5	66.9	62.7	54.4	74.2
CyberGym (Pass@1)	—	84.5	80.0	84.5	83.3	76.7	88.1
HLE with tools (Pass@1)	63.6	—	59.8	62.5	60.0	51.5	63.9
AutomationBench (Pass@1)	50.3	45.8	46.7	48.8	43.2	37.7	54.8
Agent's Last Exam (Pass@1)	28.6	26.7	27.6	28.5	25.7	25.2	31.8

The pattern is consistent. On agentic benchmarks — terminal work, software engineering, security, automation — V4.1 Flash sits at or above the top of the table, including above its own V4 Pro, and is the only open-weight model in the column above the closed frontier on several rows. On pure reasoning (GPQA Diamond) it is inside the frontier cluster but behind GPT-5.6 Sol, Opus-5.0 and K3. 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 it was released 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 and throughput (our benchmark suite, run on 2026-09-11): 65 ms average time to first token, about 106 tokens per second of average generation speed, and 88% accuracy on a 25-item factual QA set.
• 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.

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 included, prepaid balance): €0.182 per million input tokens, €0.728 per million output tokens, €0.014 per million cached input tokens.

Model	Lane	Input €/M (client)	Output €/M (client)
DeepSeek V4.1 Flash (552B, multimodal, new)	Fast access · Modal	€0.182	€0.728
DeepSeek V4 Flash 0731 (158B, previous gen)	EU sovereign · Scaleway	€0.56	€1.12

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 158B generation. Second, our cost basis for V4.1 Flash today is the public market price of the same model (0.15/0.60 USD per million at its base window), converted at the day's EUR rate and marked up like everything else on the catalog — a starting point that we will re-derive from the real serving cost of the dedicated endpoint once it carries sustained traffic.

Two practical consequences. 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. And because the dedicated endpoint can cold-start after a quiet period, the first call of a session can be slow — the following calls come back in seconds.

Sovereignty: what it actually means

On our catalog, V4.1 Flash runs on a Modal endpoint — a fast-access lane, explicitly labelled as such in the interface and in the API response (x-sovereign: false). That is the trade of serving a day-zero release before the sovereign catalogues catch up: you get the newest weights today, on a lane under non-EU control.

The weights themselves are MIT-licensed and open, so an EU-sovereign deployment of this model is technically open — it simply does not exist in our catalog yet. If EU jurisdiction is a hard requirement today, the sovereign DeepSeek in the catalog is the previous-generation V4 Flash 0731 on Scaleway. The GDPR vs CLOUD Act article explains why the operator — not the model — is what decides jurisdiction.

FAQ