Qwen 3.8 125B A6B MLX text-tower inference speedup track (Vontra/Qwen3.8-Flash-Next-MLX-4bit-MTP @ 327c8a604de613b42f84ba5e6b796c0931e8aa3b). The tower is 48 layers on a four-layer repeat: the last layer of each group (index % 4 == 3, i.e. 3/7/.../47) is full attention and the other three are gated deltanet linear attention, so 12 layers carry a KV cache and 36 carry a constant-size recurrent state. Full attention is 24 query heads over 2 KV heads at head_dim 256, partial rotary 0.25, interleaved mrope, with a QSA indexer (4 heads, 1 KV head, dim 128, budget 2048, compress 4). Every layer carries a sparse MoE block of 512 routed experts, 10 per token, moe_intermediate 640, plus a shared expert of the same width behind a shared expert gate. Hyper-connections run at width 4, lowrank 320. Layer 1 carries the n-gram / PLE table (ngram_size 3, 8 heads per n-gram, 128 split parts, 384 tensors): it is offloaded to SSD and served through a bounded LRU, not held resident. Embeddings are untied. The checkpoint also carries a vision tower, which is SKIPPED at load; the text tower is what is measured. Optimize the MLX runner, the offline transform, and the vendored MLX/Metal kernel families the forward pass dispatches. There is exactly ONE speculative arm: the NATIVE MTP head, which is EMBEDDED in the pinned target checkpoint under language_model.mtp.* (76 tensors, 1 hidden layer, hybrid full attention, riding the target's own embeddings). Nothing stages a head weight file on this track and no submission carries one: mtp-head.manifest.json is a DECLARATION only, it stays editable and optional, and it accepts source 'pinned' only -- which on this track means the head embedded in the pinned target checkpoint. Draft depth is participant-selectable and is not pinned at 1: permitted depths are 1 to 6, the non-editable engine clamps at 6 (the pinned fork serves the embedded head at depth 1 to 6 and its manifest declares depth [1, 6]), and benchd measures at depth 2 when the invocation names no depth. A request above the ceiling is clamped, not refused. The depth that actually ran is sealed per run as effective_spec, alongside the realized draft length effective_mean_draft_len. The lever is therefore to make the target and MTP forward passes and the surrounding kernel dispatch faster. The target model may never be re-quantized. The KV backend is pinned contiguous and refuses rather than degrades. Scoring is single-stream and paired, with a per-box baseline. A ranked run measures 2 pairs on one box in one job, over the one prompt the fixture names in live_golden: each pair is one serial-control leg on the organizer-staged reference tree at the fixture's baseline_reference_commit and one candidate leg at the declared draft depth. The legs run strictly one after the other and each leg loads the model once. Per role the per-token times are summed over the pairs, and composite = prefill_gain^0.25 * decode_gain^0.75 is the live ratio of those sums. Both speedup floors are 0.95 and the ceiling is 5.0, applied to that aggregate. docs/participant-contract.md section 5 is the authority.