DeepSeek V4.1 Flash Is an Architecture Reset, Not Just a Cheaper Model Despite having more total and active parameters than its predecessor, V4.1 Flash cuts working KV cache to one quarter and persistent cache storage to one eighth. Zhihu contributor Exhalation explains how it does this by deleting old modules, sharing KV states, and recomputing local context. 1️⃣ DeepSeek removed its own previous ideas 🔹 MTP: external draft models such as DSpark weakened its speculative-decoding value, while the auxiliary loss no longer justified its memory cost. 🔹 Heavily Compressed Attention: its global-summary role was ambiguous and difficult to combine with FP4 storage. 🔹 Dense warmup: V4.1 trains sparse attention from scratch rather than starting with one trillion dense-attention tokens. 2️⃣ Store less, reuse more Non-SWA KV cache moves from FP8 to FP4, while the more sensitive SWA portion remains FP8. DeepSeek no longer persists SWA cache. When a conversation forks from an earlier point, the system rebuilds only a small local window. Post-training simulated this process to limit numerical drift. A modified YOCO design provides the other major saving. Upper layers reuse the same lower-layer source representation, adding only a layer-specific projection. The result is roughly half the KV storage and close to 50% less historical prefill computation in the idealized case. 3️⃣ Sparse attention reuses its search Sparse attention lowers attention cost from O(n²) to O(kn), but finding the top-k tokens can still retain an O(n²) component. V4.1 Flash either reuses an earlier layer’s top-k result or selects a smaller candidate-block pool before re-indexing. This prevents token selection from becoming the bottleneck at long context lengths. 4️⃣ Engram and mHC were streamlined Engram replaces an expensive second-order optimizer state with a Sinkhorn-style update, removes causal convolution, and extends matching from 3-grams to 4-grams. mHC reorders residual mixing across layers, reducing estimated I/O from (4n+4)d to (3n+2)d. ✅ The larger pattern DeepSeek is not merely compressing an existing model. It is willing to discard its own previous components when a cheaper system-level design emerges. V4.1 Flash is less a smaller V4 than a new answer to one question: how much intelligence can be delivered per byte of memory and unit of inference cost? 🔗 Full analysis: #DeepSeek# #DeepSeekV41# #LLMArchitecture# #AIInfra# #KVCache# #SparseAttention#

Sparse Attention with Indexer also seems rather standard. Essentially, we use an indexer at block level in a compressed latent space. For training, the indexer is asked to predict the full attention score 1) Distill the dense attention scores into the indexer 2) KL loss on the indexer vs the full attention teacher

Long-context sparse attention has a catch: data-dependent block selection wrecks memory access kills speed. Our @MiniMax_AI M3 kernel on Blackwell answers it. KV-stationary, each block read once, ~980 TFLOP/s on a B200. See the breakdown here →

One difference between V4 and V4.1 papers is the details on sparse attention training. They say much more. We know concretely that they pretrain with 64K for 34T tokens, and then do 11T at 1M. No dense warm-up. No instabilities throughout.

Introducing SubQ - a major breakthrough in LLM intelligence. It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA), And the first frontier model with a 12 million token context window which is: - 52x faster than FlashAttention at 1MM tokens - Less than 5% the cost of Opus Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention). Only a small fraction actually matter. @subquadratic finds and focuses only on the ones that do. That's nearly 1,000x less compute and a new way for LLMs to scale.

Deepseek V4.1 Flash 552B total, 8/16B active with a new arch trained on 45T tokens, there are different active parameters for input/output tokens with the encoder/decoder arch, engram, new sparse attention, new mHC, native vision very high benchmarks (beating K3), insane efficiency, and as always amazing tech report this is probably the most novel arch i've seen in a while, pretty insane

Finally, a speedup method for Minimax H3 reference to video! ComfyUI-Ref2VA-VSA adds video sparse attention to Minimax, generating 5-second clips in ~72s on an RTX 4090. It keeps reference tokens dense while pruning video attention—2.24× faster than VDN-H3 at ~13.5GB VRAM.