MiniMax-M1

CONTEXT SCALING: null is a VERIFIED value here, checked against config rather than left blank. config.json (MiniMaxAI/MiniMax-M1) declares rope_scaling null, rope_theta 10000000, max_position_embeddings 10240000 (10M). Verified native per config; the 10M figure is the config value, not a vendor extrapolation recorded here. The first large-scale open-weight hybrid-attention reasoning model, and the origin of MiniMax's distinctive attention pattern. Vendor: 456B total / 45.9B activated; safetensors total 456.09B. config layer_types over 80 layers: 70 linear_attention + 10 full_attention - lightning attention interleaved one full layer in eight. kv_cache_bytes_per_token = 10 full-attention layers x 8 kv_heads x head_dim 128 x 2 x 2 bytes = 10 x 8 x 128 x 2 x 2 = 40,960 B/token (40 KiB). If all 80 layers were full attention it would be 80 x 8 x 128 x 2 x 2 = 327,680 B/token (320 KiB) - so the hybrid pattern buys an 8x KV reduction. ARITHMETIC CORRECTION (2026-10-03): an earlier version of this note gave 654,240 B/token (639 KiB) and a 16x factor. The printed product did not equal the printed result: 654,240 is what you get by forgetting the factor of 2 for K and V, i.e. 80 x 8 x 128 x 2 (K only) x 2 (bf16) doubled once too many times. The correct naive figure is 327,680 B/token and the correct factor is 8.00x, matching the 10-of-80 full-attention layer count. The verified kv_cache_bytes_per_token of 40,960 was never affected. MoE: 32 local experts, only 2 per token (num_experts_per_tok 2) - an unusually low top-k, and the main reason active params are 45.9B rather than something smaller. hidden 6144, intermediate 9216, vocab ~200k. rotary_dim 64 (partial RoPE). postnorm true. Context: max_position_embeddings 10240000 - an order of magnitude beyond anything else in this slice, built for very long RL/test-time-compute runs. rope_theta 10000000. M1 also ships per-layer alpha/beta scaling factors (mlp_alpha_factor, full_attn_alpha_factor, linear_attn_alpha_factor, all 3.5566) - a MiniMax-specific feature not seen elsewhere. Apache-2.0, notably more permissive than the M2 line's non-commercial terms. 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 lightning (linear) attention on 70 of 80 layers. llama.cpp's llm_arch_is_hybrid includes MINIMAX_01 and its supports_sm_tensor == false list does too, while MINIMAX_M2/M3 are in the false list but NOT in the hybrid list - consistent with the architecture change between M1 and M2 that this repo's own minimax-m2-7 record notes [[mdl2-llamacpp-arch-class-registry]]. NO ENGINE RECORD IN THIS REPO lists MiniMax-M1 by name; SGLang's generative table covers MiniMax-M2, M2.1 and M2.5 only, and TRT-LLM's only MiniMax row is MiniMaxM3SparseForConditionalGeneration [[mdl2-sglang-generative-models]] [[mdl2-trtllm-supported-models]]. So the M1 cell is UNKNOWN, not unsupported. The lightning-attention kernel requirement is a cumsum-free intra/inter-block formulation, per flop/ssm-sequential-scan-tensor-core-mismatch's workarounds and papers/lightning-attention-2 [[lightning-attention-2-tiling]].