Qwopus3.6-35B-A3B-v1-oQ6_8-mtp (oMLX build)

Local oMLX build of Jackrong/Qwopus3.6-35B-A3B-v1 — Qwen3.6 35B-A3B MoE (256 experts, 8 active ≈ ~3B active params/token), text-only (vision tower dropped). A mixed-precision quant tuned to maximize coding quality while leaving generous room for long context + OS on a 48 GB Mac.

Built 2026-05-28 on an M4 Max (48 GB). Sibling of the smaller Qwopus3.6-35B-A3B-v1-oQ4-mtp.

Quantization recipe (mixed "oQ6+8")

Faithfully re-quantized from the bf16 source — same module set quantized as the oQ4 build, but with per-category bit-widths chosen so the quality-sensitive paths stay high while the bulky MoE experts shrink:

Component	Bits	Group size
MoE experts (switch_mlp.{gate,up,down}_proj)	6	64
Shared expert (shared_expert.{gate,up,down}_proj)	6	64
Attention (self_attn.* + linear_attn.* projections)	8	64
Embeddings (model.embed_tokens, lm_head)	8	64
Router gates (mlp.gate, mlp.shared_expert_gate)	8	64
Norms / conv1d / A_log / dt_bias	bf16	—

• Effective ~6.753 bits/weight across the trunk (523 quantized modules: 512 trunk + 11 in the MTP head).
• Rationale: experts dominate the parameter count, so keeping them at 6-bit (near-fp16 quality, ~99% of benchmark accuracy) captures almost all the size savings, while the comparatively small attention + embedding tensors — the most precision-sensitive paths, especially with only 2 KV heads — are kept at 8-bit for the last sliver of fidelity.

MTP (multi-token prediction) draft head

• Qwopus's own MTP head is byte-identical to the unsloth base Qwen3.6-35B-A3B head, which drafts at 0% acceptance once quantized on this MoE (the known quantized-MTP collapse). This build reuses the distilled MTP head from Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled (stamsam) — the exact head file shipped in the oQ4 build — which is quantization-robust.
• Head precision is independent of the trunk and only affects draft speed, never output. On this higher-precision trunk it actually drafts better than on oQ4: ~89.7% acceptance (vs ~77% on the 4-bit trunk), because the trunk verifies closer to the distribution the head was distilled against.
• The draft head only proposes tokens; the Qwopus trunk verifies and accepts/rejects each one, so generated content is 100% Qwopus — the head is a pure speed feature.

Size & memory

• On disk: 28.06 GiB (30.13 GB) — 6 trunk shards (model-0000{1..6}-of-00006.safetensors)
  • model-mtp.safetensors (~0.49 GB distilled head).
• Resident in oMLX: ~27.1 GB.
• On a 48 GB Mac (default iogpu.wired_limit_mb=40960) this leaves 13 GB of wired headroom plus the rest for the OS. The hybrid architecture (only 10 of 40 layers are full-attention, 2 KV heads, head_dim 256 ≈ **20 KB/token** of KV cache; the 30 linear-attention layers carry a fixed-size recurrent state) makes long context cheap — the full 262 144-token window is comfortably reachable.

vs the oQ4 sibling

Build	Bits (effective)	On disk	Resident	MTP accept
…-oQ4-mtp	4.649 bpw	19.5 GiB	~19 GB	~77%
…-oQ6_8-mtp (this)	6.753 bpw	28.06 GiB	~27.1 GB	~89.7%

Verified

Loads in oMLX as qwen3_5_moe (batched engine); native MTP patch applied (active), MTP path activated, draft acceptance 61/68 = 89.7% on a coding prompt. Config quant block was validated tensor-by-tensor (weight_last*32 == scales_last*bits*group_size) before install — 0 mismatches across all 523 quantized modules.

Registered with mtp_enabled: true (non-default, non-pinned) in ~/.omlx/model_settings.json. Context window 262144, max_tokens 32768.

Build

Source is the local dir ~/.omlx/models/Jackrong/Qwopus3.6-35B-A3B-v1 (passed directly as hf_path). Scripts in ~/.omlx-build/: convert_qwopus_oq6_8.py → assemble_oq6_8.py → install_oq6_8.py.

Note: because the trunk default rose to 6-bit while the reused head keeps 4-bit switch_mlp/self_attn.o_proj modules, those head modules are written as explicit overrides in config.json's quantization block (they would otherwise inherit the new 6-bit default and fail to load).