Qwen3-Next 80B-A3B Instruct

ADDED 2026-10-05 to close a genuine gap: this checkpoint was already cited by name across the repo's benchmark, engine, flops and gotcha records (ev-vllm-mtp-regression-qwen3-next-80b-a3b, ctx-kv-cache-gib-at-claimed-context, ctx-rope-scaling-method-per-model, ctx-128k-concurrency-is-single-digit, gated-deltanet) while having no model record of its own, so every claim about its geometry had to live in someone else's notes. ALL GEOMETRY BELOW IS READ FROM config.json (Qwen/Qwen3-Next-80B-A3B-Instruct), NOT from the model card and NOT inferred from the name: num_hidden_layers 48, num_key_value_heads 2, head_dim 256, hidden_size 2048, num_attention_heads 16, num_experts 512, num_experts_per_tok 10, shared_expert_intermediate_size 512, full_attention_interval 4, linear_conv_kernel_dim 4, linear_key_head_dim 128, linear_num_key_heads 16, linear_num_value_heads 32, linear_value_head_dim 128, max_position_embeddings 262144, rope_scaling null, rope_theta 10000000. LAYER SPLIT: config does not enumerate the split; it gives full_attention_interval 4 over 48 layers, which resolves to attention at 0-based indices [3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47] = 12 of 48 layers, with the other 36 running gated linear attention (Gated DeltaNet) plus a short conv of kernel 4. layer_types is recorded explicitly so no reader has to re-derive it. This is the same 1-in-4 pattern the sibling record qwen3-5-35b-a3b records over its own 40 layers, which is an independent check that the rule is read the same way in both. KV CACHE DERIVED FROM CONFIG: 2 (K and V) x 12 full-attention layers x num_key_value_heads 2 x head_dim 256 x 2 bytes (bf16) = 12 x 2 x 256 x 2 x 2 = 24,576 B/token (24.0 KiB/token). At the advertised 262144 max_position_embeddings one full sequence is 262144 x 24576 = 6.0 GiB of KV, which is the point of the hybrid: an 80B-total model serving 256k context is a KV-planning problem solved by architecture rather than by tensor parallelism. For contrast at the same 80B total, GLM-4.5 (92 layers, GQA) sits at 376,832 B/token = 15.3x this figure. CONTEXT SCALING: config declares rope_scaling null with rope_theta 1e7, so the 262144 window is native - recorded native, not unknown. PARAMS: 80B total / 3B activated are the published Qwen figures; the safetensors index reports 162.65 GB of tensors, which is consistent with an 80B model whose expert weights are stored alongside the routed-expert indexing metadata rather than implying a different count. Treat 80/3 as vendor-stated.