DeepSeek-V4.1-Flash vs GLM-5

=Same as the blueprintDeepSeek-ViT, trained from scratch: 32 layers, width 1,024, 16 heads, 14-px patches embedded by a linear projection, 2D-RoPE, RMSNorm and SwiGLU. A 3×3 pixel-unshuffle cuts visual tokens 9× before a 2-layer MLP projector maps them to 5,120 dims. Images only, up to about 1344×1344 px.

=Same as the blueprintTwo Engram modules (196B parameters, FP8) at layers 1 and 14: n-gram orders 2–4, 8 hash heads per order, 2,048 dims per order, about 16M rows per head table. A sigmoid gate on each mHC stream's normalized dot product with its own key adds a shared value. The short causal conv was dropped. Rows are prefetched from host memory over RDMA.

=Same as the blueprintSingle-Pass mHC: n = 4 streams of width 5,120 and 20 Sinkhorn iterations as in V4, but each sublayer reads its input with the mixing coefficients predicted by the previous sublayer. Removing that dependency lets the Mega-mHC kernel read and write the residual once, halving activation traffic, with negligible quality loss reported.

=Same as the blueprintRMSNorm on the single 512-dim KV entry, which the report uses to bound every FP4 KV channel to about √512 ≈ 22.6, and on the 1,280-dim query latent. Unlike V4's reference code, V4.1's applies no separate per-head RMS normalization to queries.

=Same as the blueprintMLA in all 78 layers and the MTP layer, with 64 heads: 2,048-dim query latent, 512-dim KV latent plus a 64-dim rotary key (576 cached values per token per layer), q/k head dim 256 (192 + 64 rotary), v head dim 256. This is the MLA-256 variant: head dim raised from 192 to 256 with one-third fewer heads, keeping training FLOPs and parameters roughly constant while lowering decode compute.

=Same as the blueprintDSA in every layer, not alternated with MLA: a lightning indexer (32 heads × 128 dims, ReLU scores with learned per-head weights) selects the top 2,048 tokens per query and MLA attends only to those. It was added after mid-training: 1,000 warm-up steps on 14 × 202,752-token sequences, then 20B tokens of sparse adaptation. The indexer is frozen in RL and uses a deterministic torch.topk. GLM-5.2 runs it in 21 of 78 layers.

=Same as the blueprintLayers 0–1 are pure sliding-window layers; every other layer adds the 128 most recent uncompressed tokens (SWA KV kept in FP8) to its sparse selection. SWA KV is never persisted: on a prefix-cache hit it is rebuilt by replaying only the last 128 tokens (SWA Bounded Replay), with negligible measured loss.

=Same as the blueprintNo output gate. Each of the 64 heads has a learnable sink logit in the softmax denominator, per the reference inference code.

=Same as the blueprintOne MTP layer (embedding and hidden-state norms, a projection, then an MLA + DSA + MoE block) whose weights are shared across 3 prediction steps in training. Acceptance length is 2.76 vs. DeepSeek-V3.2's 2.55 at 4 speculative steps. GLM-5.2 reuses the first step's top-k indices and KV in later steps and adds rejection sampling and an end-to-end TV loss: 4.56 → 5.47 accepted tokens at 7 steps in a GLM-5.1-based ablation.

=Same as the blueprintThe release stores most weights in FP8 and routed experts in FP4. Post-training QAT extends FP4 to the main KV cache (E2M1 with one E4M3 scale per 16 channels and no global scale), indexer queries and keys stay FP4 as in V4, and the SWA KV stays FP8 because it is sensitive to quantization.