Qwopus3.6-35B-A3B-Coder-vram13-GGUF

Qwopus3.6-35B-A3B-Coder — vram13: 151 t/s coding on 18 GB of VRAM (13.5 GB, imatrix, native MTP)

A SWE-bench-62 coding fine-tune of Qwen3.6-35B-A3B, quantized to fit entirely in 18 GB of VRAM with its native MTP head intact — 151 t/s code decode on a pair of mid-range gaming GPUs.

This is the coding-tuned sibling of Qwen3.6-35B-A3B-vram13-GGUF, and the best of the family: Qwopus ships with a trained multi-token-prediction head (most GGUFs drop it), so speculative decoding uses the head that was optimized for these exact weights — no transplant, best acceptance.

File

Measured performance (RTX 3060 Ti 8 GB + RTX 3080 10 GB, ik_llama.cpp, q4_0 KV, all-VRAM)

Decode holds >100 t/s at 14.5 K tokens of depth. VRAM: ~6.6 + ~9.3 GB across the two cards.

Quality (perplexity, held-out corpus, ctx 512)

Within noise of a quant 35 % larger — the best quality-per-byte of the all-VRAM builds we tested.

Recipe

Requantized from Q8_0 with a Qwopus-specific imatrix, using ik_llama.cpp llama-quantize --custom-q (required: mainline/other forks fail on the nextn layer with missing tensor blk.N.ssm_conv1d.weight):

• Experts: q2_K middle layers 3–36, q3_K edge layers 0–2 & 37–39 (imatrix-guided)
• Attention + SSM projections: q5_K
• Shared expert, output, embeddings: q6_K
• blk.40 MTP/nextn head: preserved
• Base ftype q4_K so every custom rule fires

The principle: on small-active-expert MoE, expert weights dominate size but tolerate 2-bit with imatrix; everything on the every-token path stays 5–6-bit.

Running

Requires ik_llama.cpp for -mtp (and for loading the nextn arch reliably).

llama-server -m Qwopus3.6-35B-A3B-Coder-vram13-q2ex-imat.gguf \
  --jinja -fa on --cache-type-k q4_0 --cache-type-v q4_0 \
  --ctx-size 131072 -ngl 99 --tensor-split 44,56 --ubatch-size 256 --no-mmap \
  -mtp --ctx-size-draft 8192 --reasoning off

Notes: Qwopus is a thinking-off design — keep --reasoning off (also avoids empty-content responses). Single ≥16 GB GPU: drop --tensor-split. MTP output is verification-lossless; only speed varies with content (code accepts best).

Caveats

• 2-bit experts: expect slightly weaker edge-case reasoning than Q4-class quants — for maximum quality use a larger mixed quant with CPU offload.
• 128 K max context (ik clamps MTP models to leave room for the draft context); the non-MTP sibling repo runs 256 K.

---

Built on a Ryzen 5950X + 3060 Ti + 3080. Fine-tune by Jackrong (Apache-2.0), base model © Alibaba Cloud. Quantized with ik_llama.cpp tooling.