DeepSeek V4.1-Flash ships open weights built around a smaller KV cache
DeepSeek published V4.1-Flash open weights on Sept 10: a 552B multimodal MoE with an 890-byte KV cache per token, and vendor numbers that show where it loses.
Vlad Makarovreviewed and published
DeepSeek published open weights for DeepSeek-V4.1-Flash on Hugging Face on September 10, one day after the model appeared in the company's API as a beta whose endpoint name carried its own expiry date: deepseek-v4.1-flash-expires-on-0910. A temporary label became a permanent release within a day. The model card ships under an MIT license badge, lists 552B backbone parameters and support for contexts up to one million tokens, and announces its own thesis in the title: "Pushing the Limits of KV Cache Compression."
Read that title literally. The strongest claim in this release is about what the model costs to run, not how smart it is, and the two should not be blurred together.
A cost story disguised as a capability story
The architecture is a clear break from the V4 generation. DeepSeek-V4.1-Flash uses a Causal Encoder-Decoder design: a 40-layer transformer split into a 20-layer causal encoder followed by a 20-layer decoder. The decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. The practical effect is that the model activates only 8B parameters per token during prefill and 16B during decode, far below its 552B backbone.
Two more mechanisms carry the memory argument. SWA Bounded Replay reconstructs missing sliding-window attention states by replaying only the most recent n_win tokens, which keeps the persistent KV cache at roughly one-eighth of DeepSeek-V4-Flash's. Compressed Sparse Attention 2 assigns each attention layer one of three static modes (Full, Reindex, or Reuse), and a Hierarchical Sparse Indexer keeps the cost of deeper indexing from growing with context length. The KV cache itself is stored in FP4, with E2M1 values and one E4M3 scale per 16 channels, bringing the global footprint to 890 bytes per token. The card puts that at roughly a quarter of V4-Flash and a 437-fold reduction against V1.
| Spec | DeepSeek-V4.1-Flash | DeepSeek-V4-Flash |
|---|---|---|
| Backbone parameters | 552B | 284B |
| Activated per token | 8B prefill, 16B decode | 13B |
| Context window | up to 1M tokens | up to 1M tokens |
| Global KV cache | 890 bytes/token | roughly 4x larger (card figure) |
| License | MIT (card badge) | MIT (card badge) |
Why the cache number decides the purchase: agentic work is input-heavy. An agent reading a repository, a long tool transcript, or a million-token document pays for that input on every step, and the memory footprint of the cache is the line item that decides whether long-horizon tasks are affordable at all. DeepSeek is not arguing that V4.1-Flash out-thinks the frontier. It is arguing for a much smaller bill on the work these models already do. The rest of the release points the same direction: Engram conditional memory (196B parameters, sparsely accessed), DSpark speculative decoding, and a DeepSeek-ViT encoder trained from scratch.
The scorecard is DeepSeek's own
Every figure below is self-reported by DeepSeek, measured internally at maximum reasoning effort, which the card exposes as a continuously controllable setting from 1 to 100. Rival scores were measured by DeepSeek rather than by the rivals, and no independent replication has been published.
| Benchmark | V4.1-Flash | Opus 5.0 | GPT-5.6 Sol |
|---|---|---|---|
| Terminal-Bench 2.1 | 90.6 | 89.1 | 88.8 |
| Terminal-Bench 3.0 | 30.0 | 43.3 | 34.4 |
| Terminal-Bench 4.0 | 31.2 | 51.8 | 39.9 |
| DeepSWE v1.1 | 74.2 | 74.0 | 73.0 |
| CyberGym | 88.1 | — | 84.5 |
| AutomationBench | 54.8 | 50.3 | 45.8 |
| Agent's Last Exam | 31.8 | 28.6 | 26.7 |
| HLE (no tools) | 36.8 | 56.3 | 44.5 |
| ExploitGym | 15.3 | 22.1 | 33.7 |
| SEC-Bench Pro | 62.8 | — | 74.3 |
| ProgramBench | 20.3 | 37.0 | 23.0 |
| Codeforces (rating) | 3471 | — | — |
The table has two halves. On the cheap, high-volume tier of agentic execution, terminal work at the 2.1 difficulty, code-agent resolution, cyber-range tasks, long-horizon automation, V4.1-Flash leads or ties the field. On the harder tiers the ordering reverses sharply. Terminal-Bench 3.0 and 4.0 fall to 30.0 and 31.2 against Opus 5's 43.3 and 51.8. ExploitGym lands at 15.3 against GPT-5.6 Sol's 33.7. The model trails on SEC-Bench Pro (62.8 vs 74.3) and ProgramBench (20.3 vs 37.0), and its 36.8 on Humanity's Last Exam sits well below Opus 5's 56.3. The base-model card shows the same split: a small lead on MMLU-Pro (74.1 vs V4-Pro's 73.5) alongside declines on SimpleQA-Verified and LongBench-V2.
What the card does not say
The weights are open; the recipe is not. There is no training-corpus mixture, no description of the reinforcement-learning environments, and no data-pipeline detail beyond a general note that post-training changes came from automated synthesis of agent tasks. Pretraining is stated as 45T multimodal tokens from scratch, with context extended to 1M tokens at the 34T-token mark, and post-training as standard SFT to RL to on-policy distillation. That is a headline, not a reproducible procedure.
The API-beta throughput figures that circulated ahead of the release, 200 to 427 tokens per second, came from individual third-party testers running their own comparisons rather than a controlled benchmark, and should be read that way. During the beta the endpoint allowed 20 concurrent requests per account, and pricing was unchanged from V4-Flash.
Running it
The card is more developer-friendly than most. There is no Jinja chat template; the encoding folder carries a Python reference implementation with test cases for multi-turn chat, tool calling, and numeric reasoning effort, while deepseek-recipe offers the same format as maintained Rust libraries with Python bindings. Recommended sampling is temperature 1.0 and top_p 0.95. The model tree links 15 community quantizations and one finetune. For the pricing picture around this launch, the companion piece on DeepSeek's cost positioning takes the wider view.
The verdict
DeepSeek shipped a genuine engineering result: a million-token multimodal model whose persistent cache fits in a fraction of its predecessor's, available under a permissive license today. For teams running long-context agents on a budget, that is the number that changes the invoice. What it is not is a new top of the frontier. The vendor's own table shows the model winning the cheap, high-volume tier of agentic work and losing the hard one, and the recipe that produced those results stays inside DeepSeek. Open weights, closed method, self-graded scorecard.
