Qwen3.5-122B-A10B-Opus-Reasoning — Dynamic imatrix GGUF + MTP draft

Custom GGUFs of timteh673/Qwen3.5-122B-A10B-Opus-Reasoning — an Opus-style reasoning fine-tune of Qwen3.5-122B-A10B (a Gated-DeltaNet hybrid MoE, arch qwen35moe). This repo improves on the upstream GGUFs in two independent ways:

1. Dynamic, importance-matrix (imatrix) quantization — an XL-tier per-tensor bit allocation that keeps the sensitive tensors high. The result is better quality-per-byte than stock K-quants. Similar recipe to unsloth (hat tip).
2. A multi-token-prediction (MTP) draft for speculative decoding — a lossless +33% throughput boost (the upstream GGUFs ship without an MTP head).

I built this quant to work well with my local AI workspace tool, Familiar. Familiar is an AI-enabled workspace where you can take Notes, collaborate in wikis, chat with models, and create smart automations with local models. This model performs well for most activities that Familiar needs. Check it out here:

https://github.com/sixvolts/familiar

Derivative work. Credit for the model goes to the upstream fine-tune author and the Qwen team; the MTP draft head is grafted from Unsloth's base-model MTP GGUF (see Credits).

Files

1. Quality — perplexity (wikitext-2 test, 300 chunks @ 512 ctx, held-out)

Q4_K_XL recovers ~84% of the Q4_K_M→Q8_0 quality gap for +1 GB; Q3_K_XL matches standard Q4_K_M quality at 10 GB smaller. All from the same Q8_0 source + same imatrix, so the deltas are clean.

Functional spot-check (both quants): correct on arithmetic (17×23=391), the bat-and-ball logic trap ($0.05), strict instruction-following, factual 2-sentence answers, and iterative-code generation. Reasoning traces are coherent.

2. Speed — MTP speculative decoding (lossless)

The model carries no MTP head in the upstream GGUF, so one was grafted from Unsloth's base MTP-GGUF. It transfers to this fine-tune extremely well (the tune barely shifted the trunk's hidden states), giving a real speedup. Speculative decoding is lossless — the target model verifies every drafted token, so output is identical to running the quant alone.

n_max=2 is optimal here (a single nextn layer drafts ~2 tokens well; the 3rd costs more than it returns). Benchmarked on 4× MI100 (gfx908), one infinity fabric hive.

llama-server -m Qwen3.5-122B-A10B-Opus-Reasoning-Q4_K_XL.gguf \
             -md mtp-opus-q4kxl-draft.gguf \
             --spec-type draft-mtp -ngld 99 --spec-draft-n-max 2 \
             -ngl 99 -c 8192 -fa on --jinja

Note: always pass -c (the model's trained context is 262144; the default grabs it all and OOMs). Use a qwen35moe-capable llama.cpp (mainline supports it, incl. the GDN kernels).

Quantization recipe

• imatrix from the standard calibration_datav3 corpus (512-token chunks).
• Dynamic per-tensor types (llama-quantize --tensor-type-file):
  • Q4_K_XL: bulk experts ffn_gate/up_exps=Q4_K; ffn_down_exps=Q5_K; embeddings, output, attention, GDN ssm_*, shared experts, first/last layers=Q6_K; GDN in-proj attn_qkv=Q5_K.
  • Q3_K_XL: base Q3_K_M; experts Q3_K; ffn_down_exps/attn_qkv/first-last=Q4_K; shared experts=Q5_K; embeddings/output/attention/GDN-out=Q6_K.

Credits

• Base model: Qwen/Qwen3.5-122B-A10B (Qwen team)
• Fine-tune: timteh673/Qwen3.5-122B-A10B-Opus-Reasoning
• MTP draft head grafted from: unsloth/Qwen3.5-122B-A10B-MTP-GGUF
• Quantization, imatrix, and MTP graft: this repo. License follows the upstream model.