DeepSeek发布v4.1-Flash：新型因果编解码架构与极致推理效率

We are late to this but better than never. Have been busy finalizing the second AIE NYC, which is happening in one month. Get your tix before prices go up - we will announce speakers from Bridgewater, Ramp, Coatue, Mastercard, Vanguard, Coinbase, Blackrock, Fidelity, Point72, Capital One, JPMC, Wells Fargo, Bloomberg, A24 (yes the movie studio) Labs, Two Sigma, Apollo Global, and more next week!

The way DeepSeek pursues their research agenda is nothing short of fascinating. In between major DeepSeek versions, from v2 to v3 to v4, they have released intermediate papers with a hyperfocused architectural improvement and basically a 100% hit rate, from Math (esp GRPO), Coder, and R1, not to mention more recent work on Manifold Constrained Hyperconnections and Compressed Sparse Attention. After the enormous attention in 1H2025 from the R1 paper, DeepSeek started laying low, and for about the past year, was happy to let peers like GLM and Kimi take the lead on Open Models.

It looked dicey for a little bit, but true whalebros never wavered, and now DeepSeek are sending a weirdly mixed message by doing a completely new architecture, retiring V4 Pro and going all in on this new model, and yet only titling it v4.1 Flash, it seems to be a test of whether or not you know how to read through the basic headlines to understand true advances.

Yes, v4.1 Flash is technically behind other open models in some benchmarks. But that’s because we don’t yet have benchmarks that concisely capture what v4.1, and the broader research agenda of DeepSeek, is aiming for - the most creative and efficient use of context we have ever seen openly explained.

If you are the sort to only read model versions and benchmark headlines, you are exactly the type of superficial person that DeepSeek is looking to fool. The best way to understand DeepSeek’s enormous advance here is to look at Sebastian’s meme:

Same model name, but hardly a 0.1 bump by anyone’s standards, and they even threw in vision without making you wait for a separate model. For a better visualization you can look at all the model innovations stacked up over time from the OG encoder-decoder architecture from Attention is All You Need:

If you read our V4 Pro writeup and Engram you should be up to date on the basic architectural reading for DeepSeek as of April 2026, but what we are HUGE fans of is the prefill/decode separation introduced here, 8B in prefill (input tokens), 16B in decode (output tokens), causing our alphabet soup of “DeepSeek v4.1-Flash: 763B-P8B-D16B” if you extend the established notation for MoEs. That’s a sparsity of 1-2%, and if you read the DeepSeek v4.1 Flash tech report, combined with new tweaks like Sliding-Window Attention Bounded Replay, makes for a KV cache footprint up to 1/8 that of V4 Flash… which make it much better/faster/cheaper for long running agents:

We are so glad that DeepSeek is back publishing SOTA research. Our last highlight is their comments on post-training, where they largely seem to agree with Prof Jie Tang:

AI News for 9/9/2026-9/10/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

DeepSeek launched V4.1-Flash as a new open-weight flagship focused on extreme inference efficiency and low cost.

• Independent benchmark account Artificial Analysis reported that DeepSeek V4.1 Flash surpasses DeepSeek V4 Pro 0813 despite being much cheaper, scoring 40 on the Artificial Analysis Intelligence Index, just below GLM-5.3-Flash and above the latest V4 Pro, while being priced at $0.30 / 1M input tokens and $1.20 / 1M output tokens with cached input at $0.006 / 1M and an additional 50% off-peak discount; they also describe it as a 763B total-parameter model with 8B active input and 16B active output parameters, 1M-token context, text+image input, MIT license, and US/API availability via DeepSeek first party @ArtificialAnlys, @ArtificialAnlys, @ArtificialAnlys
• Vals called it the new #1 open-weight model on the Vals Index, ahead of Kimi K3, at just $0.30 per test, the cheapest model in the open-weight top 10; they also note the eval ran with 1M context, 384 max output tokens, temperature 1, default top-p/top-k, and high reasoning effort @ValsAI, @ValsAI, @ValsAI
• Baseten shipped day-0 support and summarized the product positioning as smarter, faster, and more efficient than DeepSeek v4 Pro 0813, with text and vision, US-only, ZDR, and 1M context @baseten
• Ollama began rolling it out to Max and Team accounts, later expanding to Pro plan subscribers @ollama, @ollama, @ollama

Architecture and paper-level technical details

The most discussed technical novelty is a causal encoder-decoder design aimed at lowering active compute and KV/cache costs.

• Artificial Analysis says the model uses a new causal Encoder–Decoder architecture, with 8B active parameters for input/prefill and 16B active parameters for output/decode @ArtificialAnlys
• Sebastian Raschka characterized V4.1 as a “big overhaul” and said they “should have called it DeepSeek V5,” explicitly highlighting the encoder-decoder setup as the key break from prior DeepSeek generations @rasbt
• Multiple technical readers reacted to the design as unusually hybrid: one called it “a very interesting mix of very conservative and sometimes old ideas in research and potentially cutting edge efficiency and hardware design in engineering” @_xjdr
• A concise architecture read from Stochastic Chasm compared the design philosophy to HySparse, NSA, and DeepSeek’s own CSA/HCA from V4, summarizing it as a local sliding-window branch plus sparse retrieval branch, suggesting this sparse/local hybrid is becoming a broader pattern @stochasticchasm
• The same account noted multimodal changes were not radical, saying DeepSeek mostly “lets the backbone handle most of it and give it visual tokens,” with 3x3 pixel unshuffle instead of the more common 2x2 @stochasticchasm
• They later flagged a “big difference from K3 on vision encoders,” implying the vision front-end diverges materially from recent Chinese peers @stochasticchasm
• TeortaxesTex observed a recurring DeepSeek pattern of doing something unusual in the first N layers—previously dense or hash-routed, now SWA-only—speculating this may reflect repeated training difficulties in early layers @teortaxesTex
• Later, the same account argued the stack is “down to 40 layers, arguably only 20 legit decoder layers,” underscoring just how aggressively DeepSeek may be compressing effective depth in decode-critical paths @teortaxesTex
• Another thread fragment from TeortaxesTex suggested DeepSeek is doing multiple compression frequencies, “it’s just all CSA2,” in response to architectural discussion around memory compression @teortaxesTex
• Nrehiew’s technical notes emphasize KV cache compression as central to the design, calling it a case study in “how obsessing over KV Cache compression gets you a hyper-efficient frontier model” @nrehiew_
• In a follow-up, nrehiew highlighted infrastructure specifics from the report: dispatch strategy to reduce long-tail stalls, router replay from previous checkpoints, management of shorter-completion off-policy effects via dataset-level capping, discard schemes, bounded off-policy ratio and loss masking, and persistent KVs and routers when a new checkpoint is updated; they also mention a final stage with full-vocab OPD on 40+ teacher models @nrehiew_
• Nrehiew concluded that the design looks cleaner than the older HSA + CSA combination in V4, saying it was “very clearly designed for inference,” and cited a striking ~890 bytes/token KV size for the benchmarked score regime @nrehiew_
• Stochastic Chasm inferred QAT for the KV cache, saying this would explain why the model performs better than peers under FP4 KV cache @stochasticchasm

Benchmark results and numbers