Qwen3.8-Flash-Next

Largest and most aggressively sparse model in the Qwen line. Vendor card explains the odd 180B safetensors total: '125B with 6B activated, plus 51B n-gram embedding and 4B MTP' - so the transformer body is 125B and the rest is a large n-gram embedding table plus multi-token prediction heads. The 51B n-gram table is not per-token active FLOPs, which is why active_params_b is 6, not 180. kv_cache_bytes_per_token = 12 full-attention layers x 2 kv_heads x head_dim 256 x 2 x 2 = 12 x 2 x 256 x 2 x 2 = 24,576 B/token (24 KiB). layer_types over 48 layers: 36 linear_attention + 12 full_attention, full_attention_interval 4, linear_conv_kernel_dim 4, linear_num_value_heads 48. Same hybrid pattern as Qwen3.5 but with a 2:3 ratio... precisely, one full attention per 4 layers. MoE: 512 experts, 10 routed per token, PLUS a shared expert (shared_expert_intermediate_size 640). hidden 2560, moe_intermediate_size 640, gqa_ratio 12 (24 heads / 2 kv heads). Context 262144. rope_parameters: rope_type default, rope_theta 10000000, partial_rotary_factor 0.25, mrope_interleaved true with mrope_section [11,11,10] - a multimodal rotary scheme (mrope) rather than plain 1-D rope, which matters for multimodal serving. LICENSE is the bespoke 'Qwen Community License 1.0', a regression from Qwen3.5's Apache-2.0 - worth checking before adopting for commercial on-prem. 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 linear attention with full attention every 4th layer, plus mrope (a multimodal rotary scheme with mrope_section [11,11,10]) and partial_rotary_factor 0.25, so it needs the linear-attention kernel AND the multimodal mrope path. llama.cpp's llm_arch_is_hybrid includes QWEN4EXP, which is this generation's arch entry, and QWEN35 [[mdl2-llamacpp-arch-class-registry]]; TRT-LLM has Qwen4ExpForCausalLM (Qwen3.8-Flash-Next, text) and Qwen3_5MoeForCausalLM rows, and ik-llama-cpp's architecture list runs to 'Qwen 3.8' [[mdl2-trtllm-supported-models]] [[ik-llama-cpp]]. TRT-LLM's multimodal matrix marks Qwen4ExpForConditionalGeneration Untested for overlap scheduler, CUDA graph, attention DP, chunked prefill, KV cache reuse and EPD disaggregated serving - every one of its cells [[mdl2-trtllm-supported-models]]. This is the clearest 'supported but unverified anywhere' cell in the repo. CONTEXT_SCALING_METHOD (native): config declares no rope_scaling with rope_theta 1e7 over max_position_embeddings 262144. NOTE the config read is Qwen/Qwen3-Next-80B-A3B-Instruct (the repo the record's own sibling cites); the 48-layer geometry and the 12-of-48 full-attention split match this record, but the checkpoint is not byte-identical. LAYER LAYOUT (qwen3-8-flash-next): config declares full_attention_interval 4 with no per-layer array, so the split is reconstructed as every 4th layer (1-based) being full attention. Matches the record's own 'X of Y' note.