DeepSeek Retired Its 1.6T Flagship for a 16B-Active Model
TL;DR
DeepSeek-V4.1-Flash landed on September 10 with MIT weights, a 552B backbone, a one-million-token context, and native image input. It activates 8B parameters per token while reading your prompt and 16B while writing the answer, and it holds its entire global KV cache in 890 bytes per token. On September 14, every API call aimed at DeepSeek-V4-Pro, the closed 1.6T flagship, gets routed here and billed at Flash rates, which is roughly a 70% cut on output. It also beats Claude Opus 5 on Terminal-Bench 2.1 and loses to it by 13 and 21 points on Terminal-Bench 3.0 and 4.0. DeepSeek printed both tables.
The part that is actually new
Everyone will lead with the benchmark win. The interesting number is the activation count.
DeepSeek-V4-Pro, the model this thing replaces, carries a 1.6T backbone and fires 49B parameters per token. V4.1-Flash carries 552B and fires 8B during prefill, 16B during decode. That is not a tuning choice, it is an architecture change, and it is the whole reason the price moved.
The trick is where the KV comes from
DeepSeek calls it a Causal Encoder-Decoder, or CED: a 40-layer transformer split into a 20-layer causal encoder followed by a 20-layer decoder. In a normal decoder-only stack, every layer computes and stores its own keys and values for the prompt, so your cache grows with depth. In CED, the decoder's global KV is projected from the final encoder hidden states instead of from each decoder layer's own hidden states.
Picture a forty-person meeting where all forty people take their own notes and keep their own notebooks. CED sends twenty people to sit through the meeting, then has them hand one shared summary to the twenty who actually have to answer questions. The answering half never rebuilds the notes.
That is also why prefill is cheaper than decode here rather than the other way around. Prompt tokens only pass through the encoder half, which per the technical report roughly halves prefill compute on long sequences. If your workload is a coding agent stuffing 400K tokens of repository into every turn, that is the bill you were actually paying.
890 bytes a token
The KV number is the one to write down. DeepSeek stacks three things to get there: Compressed Sparse Attention 2, which assigns every attention layer one of three static modes (Full, Reindex, Reuse) so layers share main KV and reuse each other's Top-K sparse indices; a hierarchical sparse indexer in the decoder that restricts later indexing layers to a candidate pool built by the first Full-mode layer; and FP4 main KV caching in E2M1 format with one E4M3 scale per 16 channels.
Result: 890 bytes per token of global KV, about a quarter of DeepSeek-V4-Flash. A separate mechanism called SWA Bounded Replay rebuilds sliding-window KV state by replaying the most recent window instead of persisting it to SSD, taking the persistent footprint to roughly one eighth of V4-Flash.
Do the multiplication and a completely full one-million-token context costs about 890 MB of KV. That is a number you can hold in your head, which has not been true of a frontier-class long-context model in a while.
It beat Opus 5. It also lost to Opus 5.
On Terminal-Bench 2.1 at maximum reasoning effort, V4.1-Flash scores 90.6 against Opus 5's 89.1 and GPT-5.6 Sol's 88.8. That is the headline, and it is real.
Two rows down in the same table, Terminal-Bench 3.0 has Opus 5 at 43.3 and V4.1-Flash at 30.0. Terminal-Bench 4.0 has Opus 5 at 51.8 and V4.1-Flash at 31.2. The harder the terminal benchmark gets, the wider the gap, in the wrong direction.
Same pattern elsewhere. On Humanity's Last Exam without tools it scores 36.8 against Opus 5's 56.3. ProgramBench: 20.3 against 37.0. NL2Repo-Bench: 64.0 against 75.3. ExploitGym: 15.3 against 22.1.
Where it wins, it wins clean. DeepSWE v1.1 resolved 74.2% against Opus 5's 74.0%. CyberGym 88.1, best in the table. AutomationBench 54.8 against 50.3. Agent's Last Exam 31.8 against 28.6. A Codeforces rating of 3471, up from V4-Pro's 3348. HLE with tools 63.9 against 63.6.
Read as a whole, that is a model tuned hard for agentic loops that is still behind on raw reasoning depth. Which, at $0.60 per million output tokens, is a trade most people will take. Worth noting that DeepSeek published the rows where it loses by 21 points. Calibrate your expectations for the rest of the industry accordingly.
The scaffold moves the number more than you want it to
Buried further down the card is a table nobody will screenshot: the same model on DeepSWE v1.1 across eight agent harnesses.
- mini-SWE: 74.2
- DeepSeek Harness Minimal: 72.6
- DeepSeek Harness Standard: 70.5
- Claude Code: 69.8
- DeepSeek Harness PTC: 67.6
- Pi: 66.2
- Codex: 65.6
- OpenCode: 65.5
That is an 8.7-point spread on identical weights, and the 74.2 everyone is quoting is the top of it. Terminal-Bench 2.1 shows the same shape: 90.6 under DeepSeek's own Minimal harness, 88.0 under Claude Code, 84.1 under Codex. If you are benchmarking this against your current stack, the scaffold you already run is worth more than the number in the announcement.
The price, and what dies for it
Per DeepSeek's pricing page, off-peak rates per million tokens:
- V4-Pro: $0.022 cached input, $0.66 uncached input, $1.98 output
- Flash: $0.003 cached input, $0.15 uncached input, $0.60 output
That is 70% off output and 77% off uncached input. Peak rates are double, and peak means 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which is the Chinese working day. If your batch job can wait until the Beijing office goes home, it halves.
The catch is what goes away. V4-Pro was API-only: 1.6T backbone, no checkpoint on Hugging Face, no public weights, ever. From September 14 its requests are served by V4.1-Flash at Flash prices. A closed trillion-scale flagship is being quietly decommissioned into an MIT-licensed model a third its size, and the open one is the survivor.
If you want to run it yourself
Weights are MIT on Hugging Face with vLLM, SGLang, and Transformers paths. Trained from scratch on 45T multimodal tokens at a 7:1 text-to-multimodal ratio, with sparse attention trained at 64K and context extended to 1M at the 34T mark. One shared expert and 384 routed experts per MoE layer, six routed experts active per token. Reasoning effort is a continuous integer from 1 to 100, so you can dial cost against accuracy per request rather than per model.
Two things to plan around. First, the disk: 552B backbone plus 196B of Engram conditional-memory parameters is roughly 748B on disk, so the 890-byte KV cache is doing nothing to make this a single-GPU model. Second, there is no Jinja chat template in this release. You get a reference Python implementation in the encoding folder, or deepseek-recipe, a set of Rust libraries with Python bindings that convert Messages, Chat Completions, and Responses API requests into DeepSeek's prompt format. Whether shipping a prompt encoder instead of a template is a courtesy or a hazing ritual depends entirely on your stack.
Key Takeaways
- DeepSeek-V4.1-Flash shipped September 10 under MIT: 552B backbone, 8B activated during prefill and 16B during decode, 1M context, native image input, 45T training tokens.
- The Causal Encoder-Decoder design projects the decoder's global KV from the final encoder hidden states, cutting global KV to 890 bytes per token, about a quarter of V4-Flash, and roughly halving prefill compute on long sequences.
- It tops Terminal-Bench 2.1 at 90.6 against Opus 5's 89.1, and loses Terminal-Bench 3.0 (30.0 to 43.3) and 4.0 (31.2 to 51.8) in the same published table.
- The same weights swing 8.7 points on DeepSWE v1.1 across eight harnesses, from 74.2 on mini-SWE down to 65.5 on OpenCode. The scaffold is part of the score.
- From September 14, V4-Pro API traffic routes to V4.1-Flash at Flash rates: $0.15 uncached input and $0.60 output per million off-peak, roughly 70% off V4-Pro output.
- Self-hosting means about 748B parameters on disk once the 196B Engram modules are counted, and no Jinja chat template in the box.
Sources: DeepSeek-V4.1-Flash model card on Hugging Face, DeepSeek-V4.1-Flash technical report, DeepSeek API models and pricing, SiliconANGLE, VentureBeat, MarkTechPost