Kimi-Linear-48B-A3B-Instruct

Hybrid KDA + MLA, and the best KV-per-active-param ratio in the whole slice. Vendor: 48B total / 3B activated; safetensors total 49.12B. config linear_attn_config enumerates 27 layers split into 7 full_attn_layers [4,8,12,16,20,24,27] and 20 kda_layers. Only the 7 full layers cache anything. kv_cache_bytes_per_token = 7 layers x (kv_lora_rank 512 + qk_rope_head_dim 64) x 2 bytes = 7 x 576 x 2 = 8,064 B/token (7.9 KiB). That is 8.7x smaller than DeepSeek-V3's 68.6 KiB at a comparable 3B active, and 46x smaller than GLM-4.5's 368 KiB. At 128k context a full sequence is 0.98 GiB of KV. The 20 KDA layers hold a constant-size recurrent state, which is why the number barely moves with context length. Note this is MLA on the full-attention layers (kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, mlp head_dim 72 for the linear part). q_lora_rank is null so the query path is not low-rank here. MoE: 256 experts, 8 per token, PLUS 1 shared expert, first_k_dense_replace 1, moe_intermediate_size 1024, hidden 2304 (narrow), intermediate 9216. routed_scaling_factor 2.446. Context 1048576 (model_max_length). rope_theta 10000, no rope_scaling. MIT. Good local fit for the kind of setup that would struggle with a 671B. SERVICEABILITY JOIN (added 2026-10-03; the architecture x engine matrix this record was missing, with a citation per cell in docs/13-model-compatibility.md): FAMILY=hybrid-attention-ssm. REQUIRED KERNEL CLASS: two kernel families in one graph - a chunked linear-attention / delta-rule kernel for the SSM or linear-attention layers, and an ordinary or MLA decode kernel for the full-attention minority - plus a fixed-size recurrent state allocator. Three independent records say the recurrent branch is the hard part: flop/ssm-recurrent-state-update (the state is a fixed per-sequence tax that never amortises, so it scales with concurrency and NOT with context); flop/ssm-sequential-scan-tensor-core-mismatch (the recurrence is a multiplicative chain that tensor cores cannot express, so it must run as a sequential scan; the published SSD-vs-fused-scan gap is 2-8x for identical mathematics); and llama.cpp's own dispatch table, which puts every hybrid and SSM architecture in its supports_sm_tensor == false list [[mdl2-llamacpp-arch-class-registry]] - the engine agreeing with the arithmetic. That same list also shows llama.cpp needed ONE shared hybrid code path (llm_arch_is_hybrid, 18 entries: JAMBA, FALCON_H1, PLAMO2, GRANITE_HYBRID, LFM2, LFM2MOE, NEMOTRON_H, NEMOTRON_H_MOE, QWEN3NEXT, KIMI_LINEAR, BAILINGMOE3, KIMI_K3, GLM5_NEXT, QWEN35, QWEN35MOE, QWEN4EXP, DEEPSEEK4, MINIMAX_01) covering exactly the families vLLM and SGLang each ship as separate rows, which is why hybrid support arrives per engine release rather than at once. SUPPORTED BY (cited): vLLM native rows MambaForCausalLM, Mamba2ForCausalLM, JambaForCausalLM, FalconMambaForCausalLM, FalconH1ForCausalLM, NemotronHForCausalLM, GraniteMoeHybridForCausalLM, Qwen3NextForCausalLM, KimiLinearForCausalLM, Lfm2ForCausalLM, Lfm2MoeForCausalLM, OlmoHybridForCausalLM, KimiK3ForConditionalGeneration, BailingMoeV3ForCausalLM, Qwen3_5 rows [[mdl2-vllm-supported-models-2026-10]]; SGLang generative table names Granite 4.0 micro as 'hybrid Mamba-MoE ... combining gated short convolutions with a small number of grouped query attention blocks', Kimi Linear as 'hybrid linear attention model ... Kimi Delta Attention (KDA) for up to 6x faster decoding and 75% KV cache reduction vs full attention', Nemotron Nano as 'hybrid Mamba-Transformer architecture', Ling as 'hybrid MoE model: per-layer mix of DeepSeek-style MLA full attention and Qwen3 Gated-Delta-Net (GDN) linear attention', LFM2 as 'hybrid backbone', Mamba-Codestral as 'Mamba2 ... holds no KV cache', and its multimodal page carries the descriptor 'hybrid KDA/MLA' [[mdl2-sglang-generative-models]] [[mdl2-sglang-multimodal-language-models]]. TensorRT-LLM has rows for NemotronHForCausalLM, Qwen3NextForCausalLM, Qwen3_5MoeForCausalLM, Qwen4ExpForCausalLM, KimiLinearForCausalLM and KimiK3ForConditionalGeneration but NO row at all for MambaForCausalLM, Mamba2ForCausalLM, FalconH1ForCausalLM, Lfm2ForCausalLM or GraniteMoeHybridForCausalLM [[mdl2-trtllm-supported-models]]. PER-FAMILY FEATURE TRAPS, all from the TRT-LLM key-models matrix, which is the only per-architecture feature matrix in this repo [[mdl2-trtllm-supported-models]]: Qwen3NextForCausalLM has Sliding Window Attention = No and Speculative Decoding = No; Kimi K3 has no MTP or EAGLE-3 head at all (spec method DSpark only) and needs Blackwell SM100; MiniMax-M3 sparse-attention path supports neither KV cache reuse nor MTP in this release; DeepSeek-V4 and Step-3.7 are 'Untested' for disaggregated serving. KERNEL BUGS THAT ARE ARCHITECTURE-SPECIFIC AND REPRODUCED: (1) vLLM's EAGLE last-block drop (drop_eagle_block) costs 6-10 points of prefix-cache hit rate, and on a hybrid GDN/Mamba layout the scheduler alignment unit is 1,648 tokens instead of a 16-token dense block, so the same protection costs ~100x more cache hit - 97.5% to 86.9-91.7% hit rate and 30-40% batch throughput on prefix-reusing workloads [[eagle-prefix-cache-last-block-drop]]. (2) llama.cpp decode collapses on hybrid gated-deltanet GGUF models at high KV position: 33 t/s at 68K to 1.4 t/s at 91K with the model fully resident, across KV quant types and with FA on or off, reproduced on Windows and WSL2 on master and release [[gated-deltanet-decode-collapse-long-context]]. (3) llama.cpp's SYCL path has a regression specific to hybrid linear-attention graphs - qwen3next/qwen3_35 architecture models crash in ggml_sycl_op_mul_mat or return empty output on 2x Arc Pro B60, bracketed between builds b8477 (works) and b9479 (fails), while dense-attention models on the same cards are unaffected [[sycl-hybrid-linear-attention-arch-crash-intel-arc]]. (4) FlashInfer rejects large head sizes at engine init for the head geometry these hybrids introduce [[flashinfer-rejects-large-head-dim]]. Hybrid MLA + KDA again: needs the MLA kernel for its 7 full-attention layers AND the delta-rule kernel for 20 KDA layers. vLLM has a dedicated KimiLinearForCausalLM row with PP supported; TRT-LLM's row exists but the footnote says Kimi K3 is Blackwell SM100 family only - that footnote is attached to KimiLinearForCausalLM too, so treat the Blackwell restriction as applying to this row [[mdl2-trtllm-supported-models]] [[mdl2-vllm-supported-models-2026-10]]. CONTEXT_SCALING_METHOD (none-verified): config declares rope_scaling null with rope_theta 10000 while the record cites 1048576 context; no scaling declaration exists to classify. LAYER LAYOUT (kimi-linear-48b-a3b): config has no layer_types array; the split is reconstructed from linear_attn_config.full_attn_layers vs .kda_layers. Those two lists leave gaps - layer 0 (and, for Kimi K3, index 92-vs-93 off-by-one region) are short-conv mixer layers that are neither full attention nor KDA, recorded here as 'short_conv' rather than silently dropped.