The X account @stochasm — an inference and systems engineer with a technical following — spent yesterday morning posting a live read-through of DeepSeek’s V4.1-Flash tech report. It is 27 posts long, one observation at a time, with no tidy summary at the end, and it stops mid-pass at “gonna take a brunch break.” Which is exactly why it is useful: you get the reasoning as he encounters each design choice rather than a conclusion somebody else smoothed out. Here it is, with his judgments attributed to him and his uncertainty left in.

The headline architecture. The model is a causal encoder-decoder. He finds the name confusing — it is not really decoder-decoder, he argues, because you never use the first half to decode latents into anything — while conceding it is “kinda accurate.” Everyone compared the design to YOCO, which calls the same arrangement decoder-decoder, so he half-suggests just using that word and moves on (“not that this is important either way”).

The part with commercial consequences. Benchmarks are, in his word, “pretty crazy” — and the KV cache is roughly four times smaller than V4-Flash. He reads the whole design as inference-informed rather than capability-driven:

  • Cheap prefill, small KV cache, and a layout he expects to be friendly to prefill/decode disaggregation, with good utilization at large-batch prefill
  • His summary line: “inference efficiency is now the sole purpose of architectural modifications, in a sense”
  • FP4 KV cache with no apparent degradation, on top of CSA compression — the model “barely uses any KV cache at all”
  • On persistence, a nice piece of practical reasoning: within a single request the cache has to live in HBM or RAM, but between requests, if it costs a lot of space, why not simply re-prefill? At 500K context that is a negligible amount of compute

Design lineage and the choices he questions. He files it next to hysparse, NSA, and DeepSeek’s own CSA/HCA from V4 — a local sliding-window branch plus a sparse retrieval branch, an approach he says is “solidifying as a broad view of things.” Two things he flags as unexplained: why only the first two of roughly twenty layers are sliding-window-only (two layers saves little on prefill, though less-global is cheaper, so he suspects a story behind it), and why V4 bothered with two alternating compression frequencies at all — he prefers this release’s “pure CSA2.” He also notes that the design breaks DeepSeek’s own “v = new pretrain” naming convention, which surprised him, and reads the model as a simplification of V4 that they claim trained more stably.

Mechanical detail, if you care about the internals. In the causal encoder-decoder, the decoder does not attend to the encoder’s per-layer KVs. It periodically reads the final encoder states — which he argues makes sense, those being the richest — and it is not cross-attention, just self-attention with prefilled KVs. Cross-attention would attend only to the encoder’s KVs, which he assumes they ablated. It reminded him of MoDA, “an allow-KVs-from-earlier-layers type of approach.” Separately: Engram made it into the model alongside an mHC simplification, and this engram is the largest he has seen — qwen-3.8-flash-next carried 51B n-gram parameters. Sparse attention arriving without a dense-attention warmup he takes as a good sign.

Multimodal and Muon, briefly. Modality-specific experts are there, with a load-balancing term over text and image tokens. He asks whether you actually want those experts to emerge or want them discouraged for more embedding overlap between modalities, and says he does not see why modality imbalance is harmful unless it is extreme enough that you are effectively stuck with a smaller model. Otherwise the multimodal introduction is unremarkable: let the backbone carry it, hand it visual tokens, and note that a 3x3 pixel unshuffle is more aggressive than the 2x2 others use. On the Muon choice, he finds removing convolutions an interesting motivation, given Keller’s original post said Muon works with convolutions via flattening — so maybe they tried it and did not like the result.

What to weigh. This is one practitioner reading in public, at speed, and guessing in places — he says so himself, and he has not finished the report. But the shape of the thing is clear from his pass: every architectural decision in this release is aimed at what it costs to serve, which is precisely what a reader feels as price and latency. He ends with a note that no critics are in the recipe yet, “though i feel like it’s only a matter of time until they pull out some value modeling.”