Qwen3 32B

ADDED 2026-10-05 to give the repo the dense 32B point of the Qwen3 line, which was missing between the 8B (qwen3-8b) and the MoE 30B-A3B / 235B-A22B (qwen3-30b-a3b, qwen3-235b-a22b). Without it, the only dense Qwen3 serving reference in the repo is an 8B model, which is the wrong size for most single-80GB-GPU deployments people actually run. GEOMETRY DERIVED FROM config.json (Qwen/Qwen3-32B): num_hidden_layers 64, num_attention_heads 64, num_key_value_heads 8, head_dim 128 (explicit, not derived), hidden_size 5120, intermediate_size 25600, max_position_embeddings 40960, rope_scaling null, rope_theta 1000000, sliding_window null. KV CACHE FROM CONFIG: 2 (K and V) x 64 layers x 8 KV heads x 128 head_dim x 2 bytes (bf16) = 2 x 64 x 8 x 128 x 2 = 262,144 B/token (256.0 KiB/token). At 40960 max_position_embeddings one full sequence is 40960 x 262144 = 10.0 GiB of KV - which is the practically important number here and the reason this record was worth adding: a dense 32B fits on one 80GB card in bf16 weights (~64 GB) but only barely leaves room for KV, so it is a short-context or low-concurrency model on that hardware, not a 32k-context model. Compare the MoE sibling qwen3-30b-a3b at 48 layers x 4 KV heads x 128 x 2 x 2 = 98,304 B/token, i.e. the 32B costs 2.67x the KV per token of the 30B-A3B despite similar total size - MoE buys KV headroom as well as FLOP headroom. CONTEXT SCALING: config declares rope_scaling null, so the 40960 window is native - recorded native. NOTE the 40960 window is unusually small for a 2025 32B open-weight model and is a real limit of this checkpoint, not a display convention; anyone planning around the Qwen3 '256k' marketing number is planning around a different model. PARAMS: 32B is the published size; safetensors index reports 65.52 GB of tensors, consistent with bf16 weights plus an untied embedding table.