Skip to content
archdiff

models / compare / kimi-k3-vs-qwen3-8

Kimi K3 vs Qwen3.8-2.4T-A95B

How Kimi K3 (Moonshot AI, July 2026) and Qwen3.8-2.4T-A95B (Alibaba Qwen, August 2026) are built, drawn on the same November 2026 blueprint. They make different choices on 9 of 27 components. Select a component in either diagram to find it in the table.

--- frontier-llm@2026-11+++ kimi-k3@2026-07+0~5−3

Kimi K3

Moonshot AI, July 2026, 2.8T total, 104B active

Text · code · tool outputText · code · tool outputImages · video · audioImages · video · audioByte-level BPE tokenizer — 128K–262K vocabByte-level BPE tokenizer128K–262K vocabVision / audio encoder — native ViT or encoder-free patchesVision / audio encodernative ViT or encoder-free patchesToken embedding — media embeds spliced in placeToken embeddingmedia embeds spliced in placeEngram n-gram memory — hashed 2–4-gram lookup (V4.1)Engram n-gram memoryhashed 2–4-gram lookup (V4.1)Block Attention Residuals (AttnRes) — Residual & normBlock Attention Residuals (AttnRes)Residual & normRMSNorm — pre-normRMSNormpre-normQK-Clip and normalized latents (no per-head QK-Norm) — Residual & normQK-Clip and normalized latents (no per…Residual & normPartial RoPE / NoPE — rotate 64 dims, or nonePartial RoPE / NoPErotate 64 dims, or noneGated DeltaNet / KDA — linear · ~3 of every 4 layersGated DeltaNet / KDAlinear · ~3 of every 4 layersLatent attention (MLA) — global · GQA in many modelsLatent attention (MLA)global · GQA in many modelsSparse / compressed attn — DSA indexer · CSA / HCASparse / compressed attnDSA indexer · CSA / HCASliding-window attention — local · 128–1,024-token windowSliding-window attentionlocal · 128–1,024-token windowOutput gate + sink — σ(x·W_g) ⊙ o · learnable sinkOutput gate + sinkσ(x·W_g) ⊙ o · learnable sink⊕ residual⊕residualRMSNorm — pre-normRMSNormpre-normSigmoid router with Quantile Balancing — Mixture of expertsSigmoid router with Quantile BalancingMixture of expertsStable LatentMoE routed experts (SiTU-GLU) — Mixture of expertsStable LatentMoE routed experts (SiTU-…Mixture of expertsShared expert — 1–2 · always activeShared expert1–2 · always active⊕ residual⊕residualFinal RMSNormFinal RMSNormLM head — untied · vocab logitsLM headuntied · vocab logitsMTP layer turned EAGLE-3 drafter (not in open weights) — Input & outputMTP layer turned EAGLE-3 drafter (not…Input & outputNext-token logitsNext-token logitsDraft tokensDraft tokensMuon optimizer — Newton–Schulz · QK-Clip · per-headMuon optimizerNewton–Schulz · QK-Clip · per-headFP8 → FP4 training — block-scaled FP8 · MXFP4/NVFP4 QATFP8 → FP4 trainingblock-scaled FP8 · MXFP4/NVFP4 QATLong-context extension — staged 4K → 1M · YaRNLong-context extensionstaged 4K → 1M · YaRNRL with verifiable rewards — GRPO family · async agentic RLRL with verifiable rewardsGRPO family · async agentic RLOn-policy distillation — reverse KL to specialist teachersOn-policy distillationreverse KL to specialist teachersSpeculative decoding — MTP self-drafting · DSparkSpeculative decodingMTP self-drafting · DSparkKV-cache compression — FP4 KV · cross-layer reuse · SSDKV-cache compressionFP4 KV · cross-layer reuse · SSD
AttentionMixture of expertsResidual & normInput & outputTraining & inferenceData flowResidual / skipAddedReplacedNot used

--- frontier-llm@2026-11+++ qwen3-8@2026-08+0~3−4

Qwen3.8-2.4T-A95B

Alibaba Qwen, August 2026, 2.4T total, 95B active

Text · code · tool outputText · code · tool outputImages · video · audioImages · video · audioByte-level BPE tokenizer — 128K–262K vocabByte-level BPE tokenizer128K–262K vocabVision / audio encoder — native ViT or encoder-free patchesVision / audio encodernative ViT or encoder-free patchesToken embedding — media embeds spliced in placeToken embeddingmedia embeds spliced in placeEngram n-gram memory — hashed 2–4-gram lookup (V4.1)Engram n-gram memoryhashed 2–4-gram lookup (V4.1)Single residual stream — Residual & normSingle residual streamResidual & normRMSNorm — pre-normRMSNormpre-normQK-Norm — per-head RMSNorm on q, kQK-Normper-head RMSNorm on q, kPartial RoPE / NoPE — rotate 64 dims, or nonePartial RoPE / NoPErotate 64 dims, or noneGated DeltaNet / KDA — linear · ~3 of every 4 layersGated DeltaNet / KDAlinear · ~3 of every 4 layersGated grouped-query attention — AttentionGated grouped-query attentionAttentionSparse / compressed attn — DSA indexer · CSA / HCASparse / compressed attnDSA indexer · CSA / HCASliding-window attention — local · 128–1,024-token windowSliding-window attentionlocal · 128–1,024-token windowOutput gate + sink — σ(x·W_g) ⊙ o · learnable sinkOutput gate + sinkσ(x·W_g) ⊙ o · learnable sink⊕ residual⊕residualRMSNorm — pre-normRMSNormpre-normSoftmax top-k router — Mixture of expertsSoftmax top-k routerMixture of expertsRouted experts — top-6–22 of 256–896 · SwiGLURouted expertstop-6–22 of 256–896 · SwiGLUShared expert — 1–2 · always activeShared expert1–2 · always active⊕ residual⊕residualFinal RMSNormFinal RMSNormLM head — untied · vocab logitsLM headuntied · vocab logitsMTP module(s) — predict t+2 … t+kMTP module(s)predict t+2 … t+kNext-token logitsNext-token logitsDraft tokensDraft tokensMuon optimizer — Newton–Schulz · QK-Clip · per-headMuon optimizerNewton–Schulz · QK-Clip · per-headFP8 → FP4 training — block-scaled FP8 · MXFP4/NVFP4 QATFP8 → FP4 trainingblock-scaled FP8 · MXFP4/NVFP4 QATLong-context extension — staged 4K → 1M · YaRNLong-context extensionstaged 4K → 1M · YaRNRL with verifiable rewards — GRPO family · async agentic RLRL with verifiable rewardsGRPO family · async agentic RLOn-policy distillation — reverse KL to specialist teachersOn-policy distillationreverse KL to specialist teachersSpeculative decoding — MTP self-drafting · DSparkSpeculative decodingMTP self-drafting · DSparkKV-cache compression — FP4 KV · cross-layer reuse · SSDKV-cache compressionFP4 KV · cross-layer reuse · SSD
AttentionMixture of expertsResidual & normInput & outputTraining & inferenceData flowResidual / skipAddedReplacedNot usedNot disclosed

Component by component

Kimi K3 and Qwen3.8-2.4T-A95B differ on 9 of 27 components.

Blueprint componentKimi K3Qwen3.8-2.4T-A95B

Same as the blueprintMoonViT-V2: a 27-layer, 401M-parameter ViT (14-px patches, width 1,024, 12 heads) with RMSNorm and no bias terms, trained from scratch with the language-model loss instead of a SigLIP initialization. Images and video share weights (factorized spatial/temporal attention, temporal pooling); a 2×2 pixel shuffle cuts tokens 4× before an MLP projector, allowing inputs up to 3,584 × 3,584 px.

Not used

Block Attention Residuals (AttnRes)Each sublayer reads a softmax-weighted mix of the embedding and earlier 12-layer block summaries, not one running sum.

Single residual streamOne d = 8,192 pre-norm residual stream with zero-centred RMSNorm gains; no hyper-connections or depth attention.

QK-Clip and normalized latents (no per-head QK-Norm)No per-head RMSNorm on q and k: MLA logits are held by K2's weight clipping, KDA's q and k are L2-normalized.

Same as the blueprintPer-head RMSNorm (zero-centred gain) over the 256-dim heads on q and k in the 23 gated-attention layers, applied before RoPE. The Gated DeltaNet layers instead L2-normalize q and k inside the kernel.

Same as the blueprintGated MLA in 24 layers: every fourth layer (4, 8, …, 92) plus layer 93. 96 heads; queries via a 1,536-dim latent, keys and values from a 512-dim latent plus a 64-dim shared key slot (576 cached values per token per layer); q/k head dim 192, v head dim 128. The attention output is kept in FP32 during training to avoid flash attention's biased rounding.

Gated grouped-query attentionEvery fourth layer: full causal GQA with 64 query heads sharing 4 KV heads of dim 256, QK-Norm and a sigmoid output gate.

Sigmoid router with Quantile BalancingSigmoid scores pick 16 of 896 experts; each expert's selection bias is reset every step from a quantile of routing margins.

Softmax top-k routerA linear router scores 512 experts; softmax, keep the top 10, renormalize. No selection bias.

Stable LatentMoE routed experts (SiTU-GLU)896 SiTU-GLU experts run in a 3,584-dim latent (half of d); 16 fire per token; an RMSNorm precedes the up-projection.

Same as the blueprint512 SwiGLU experts per layer with intermediate size 2,048 (twice Qwen3.5's) at the full width 8,192, stored as a fused [512, 4,096, 8,192] gate/up tensor; 10 are active per token, about 2% of the pool. Routed experts hold 2.37T of the 2.45T parameters and are the only weights converted in the official FP8 checkpoint.

MTP layer turned EAGLE-3 drafter (not in open weights)One pretrained MTP layer is fine-tuned into an EAGLE-3-style draft model that reads features from three depths.

Same as the blueprintOne MTP layer (~26.4B parameters): a full gated-attention + 512-expert MoE block. Its input is a linear projection (16,384 → 8,192) of the RMSNorm'd next-token embedding concatenated with the RMSNorm'd trunk state. It reuses the main embedding and LM head (mtp_use_dedicated_embeddings: false). Per the model card it was trained with multiple steps, and vLLM applies the one layer repeatedly to draft several tokens.

Same as the blueprintPer-Head Muon: momentum of the Q, K and V projections is split along the head dimension and each head's block is orthogonalized separately, equalizing update scale across heads; other matrices use standard Muon. Pretraining also keeps Kimi K2's weight-clipping mechanism, a cosine LR schedule with 1% linear warmup and weight decay 0.1.

Not disclosed

Same as the blueprintMulti-teacher on-policy distillation merges the nine domain × effort experts: for a sampled domain and effort level, the matching teacher scores each student token with the clipped log-ratio clip(sg(log π_teacher − log π_θ), −R_max, R_max), used as a dense reward inside the RL framework. Top-k distillation objectives showed no advantage.

Not disclosed

More comparisons