How to use from
Pi
Start the llama.cpp server
# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf localweights/Qwen3.6-35B-A3B-MTP-IMAT-IQ4_XS-Q8nextn-GGUF:IQ4_XS
Configure the model in Pi
# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:9931/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "localweights/Qwen3.6-35B-A3B-MTP-IMAT-IQ4_XS-Q8nextn-GGUF:IQ4_XS"
        }
      ]
    }
  }
}
Run Pi
# Start Pi in your project directory:
pi
Quick Links

Qwen3.6-35B-A3B-MTP-IMAT-IQ4_XS-Q8nextn (GGUF)

35B MoE Qwen3.6-A3B trunk (3B active) + embedded NextN-MTP head, quantized for single-GPU inference.

  • Trunk: IQ4_XS (imatrix-calibrated)
  • MTP head: Q8_0 (NextN, kv_only_nextn=true)
  • File size: ~18.3 GB
  • Min VRAM: ~21 GB at 32K ctx with KV Q8/Q8

The MTP head is embedded inside the GGUF β€” no separate drafter file. Recent llama.cpp builds activate it via --spec-type draft-mtp.

Quick start (no spec decode)

llama-server -m Qwen3.6-35B-A3B-MTP-IMAT-IQ4_XS-Q8nextn.gguf \
  -ngl 999 -fa on -c 32768 --parallel 1 \
  -ctk q8_0 -ctv q8_0 --kv-unified \
  --host 0.0.0.0 --port 8080 --jinja

With speculative decoding (recommended)

Requires llama.cpp built from master after ggml-org/llama.cpp#22673:

llama-server -m Qwen3.6-35B-A3B-MTP-IMAT-IQ4_XS-Q8nextn.gguf \
  -ngl 999 -fa on -c 32768 --parallel 1 \
  -ctk q8_0 -ctv q8_0 --kv-unified \
  --spec-type draft-mtp \
  --spec-draft-n-max 2 \
  --spec-draft-p-min 0.0 \
  --host 0.0.0.0 --port 8080 --jinja

For A3B MoE, smaller --spec-draft-n-max wins β€” accept rate decays faster with depth than dense models. n_max=2 is the sweet spot. For MoE at deeper chains (n>=3), --spec-draft-p-min 0.0 outperforms p_min=0.75 because per-step sampler overhead from top_k+softmax exceeds the accept-rate savings on this arch.

Benchmarks

Single-stream decode on a 24 GB consumer GPU (RTX 3090 Ti), CUDA build, full-GPU offload, FA on, N=512 generated tokens, 5-warm-run mean.

Production config (ctx=32K, KV Q8/Q8)

config decode tok/s
no spec (greedy) ~120 (varies)
--spec-type draft-mtp --spec-draft-n-max 2 --spec-draft-p-min 0.0 ~200 (+67%)

Smaller ctx variant (ctx=8192, KV q4_0/q4_0, p_min=0)

n_max tok/s
1 189.5
2 198.7 (+24%)
3 191.2
4 167.6

n_max=2 wins; deeper chains regress (MoE accept rate falls faster with depth).

Quant details

  • Trunk weights: IQ4_XS calibrated with an imatrix derived from a mixed-domain calibration set.
  • MTP head: kept at Q8_0 (NextN). Sensitive to precision.
  • KV cache: Q8_0/Q8_0 recommended.

llama.cpp requirements

  • Build supporting Qwen3.6 NextN MTP (LLM_ARCH_QWEN35MOE_MTP). Merged upstream in #22673.
  • --spec-draft-p-min 0.75 requires recent llama.cpp that honors p_min on the DRAFT_MTP path.

Sister models

Reference

Apache 2.0 license.

Downloads last month
398
GGUF
Model size
36B params
Architecture
qwen35moe
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for localweights/Qwen3.6-35B-A3B-MTP-IMAT-IQ4_XS-Q8nextn-GGUF

Quantized
(1,003)
this model

Collection including localweights/Qwen3.6-35B-A3B-MTP-IMAT-IQ4_XS-Q8nextn-GGUF