Schestex/Qwen3.8-27B-ColdFusion-NVFP4-NInfer

Qwen3.8-27B ColdFusion NVFP4 NInfer

NInfer NVFP4 artifact of DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1, calibrated with a direct agentic coding dataset mix.

The goal of this build is local agentic coding, tool calling, repository work and long agent loops on NVIDIA Blackwell GPUs.

Artifact variants

This repository provides two NInfer artifacts:

Full artifact

qwen3_8_27b_coldfusion_nvfp4_direct-s256-seed38027.ninfer

Includes:

• Text
• MTP
• Vision
• DFlash2

No-Vision artifact

qwen3_8_27b_coldfusion_nvfp4_novision-s256-seed38027.ninfer

Includes:

• Text
• MTP
• DFlash2
• No Vision weights

Calibration

Direct agentic calibration was built from three pinned datasets:

Source	Samples
nvidia/Open-SWE-Traces	160
thoughtworks/agentic-coding-trajectories	64
nebius/SWE-agent-trajectories	32
Total	256

Category mix

Category	Samples
TypeScript / Svelte	48
Go	36
DevOps / Git / CI	32
C++	24
Kotlin / Gradle	24
JVM / Gradle	16
Svelte / TypeScript	16
Resolved SWE trajectories	16
Recovery trajectories	16
General agentic	16
Long agent loops	12

Dataset revisions

nvidia/Open-SWE-Traces@31cfd32021f674a1bbd5ff9f56a2151436fe2be3
thoughtworks/agentic-coding-trajectories@cef72d1f4d0caabf85937adf8337a14b7522c782
nebius/SWE-agent-trajectories@68195a1450865274106246d0d0296a1d6807b88e

Build reproducibility

llm-compressor:
b52e76d66a6f47275c33dd59342a90ae7c50d34a

NInfer:
d49296868dcc17bd478ec185f0d3a801bcc0bf56

Calibration:
256 samples
sequence length 8192
seed 38027

The repository also includes the converter report:

qwen3_8_27b_coldfusion_nvfp4_direct-s256-seed38027.ninfer.conversion.json

Verify the artifact

sha256sum qwen3_8_27b_coldfusion_nvfp4_direct-s256-seed38027.ninfer

Expected:

ae61476e16c641adcaf101e34001cff1fc3456f068083fc5e4eade8a0bcb6d9f  qwen3_8_27b_coldfusion_nvfp4_direct-s256-seed38027.ninfer

Run with NInfer Lab

Example for a 24 GB Blackwell GPU:

./ninfer-lab serve \
  --build latest \
  --model /opt/llm-models/qwen3_8_27b_coldfusion_nvfp4_direct-s256-seed38027.ninfer \
  --gpu 0 \
  --port 8080 \
  --detach \
  -- \
  --host 0.0.0.0 \
  --port 8080 \
  --max-context 32768 \
  --kv-capacity 32768 \
  --max-concurrency 1 \
  --kv-dtype int8

For larger context windows, --kv-capacity auto is recommended when VRAM is tight:

./ninfer-lab serve \
  --build latest \
  --model /opt/llm-models/qwen3_8_27b_coldfusion_nvfp4_direct-s256-seed38027.ninfer \
  --gpu 0 \
  --port 8080 \
  --detach \
  -- \
  --host 0.0.0.0 \
  --port 8080 \
  --max-context 123904 \
  --kv-capacity auto \
  --max-concurrency 1 \
  --kv-dtype int8

On an RTX PRO 4000 Blackwell SFF 24 GB, a fixed 123904 INT8 KV capacity was slightly above the available runtime reservation for this artifact, while a smaller KV pool works.

Smoke test

curl -s http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "qwen3.8-27b",
    "messages": [
      {
        "role": "user",
        "content": "Reply with exactly: COLD_FUSION_OK"
      }
    ],
    "temperature": 0,
    "max_tokens": 32
  }'

Download

Download the repository with the Hugging Face CLI:

hf download Schestex/Qwen3.8-27B-ColdFusion-NVFP4-NInfer \
  --local-dir ./Qwen3.8-27B-ColdFusion-NVFP4-NInfer

Or download the NInfer artifact directly:

hf download Schestex/Qwen3.8-27B-ColdFusion-NVFP4-NInfer \
  qwen3_8_27b_coldfusion_nvfp4_direct-s256-seed38027.ninfer

Notes

• This repository contains a NInfer artifact, not a standard Transformers checkpoint.
• The build is targeted at NVIDIA Blackwell hardware and agentic coding workloads.
• Calibration is post-training quantization calibration; the source model was not fine-tuned on the calibration samples by this build process.
• Performance and quality should be evaluated against the original model and other NInfer artifacts for the intended workload.

Credits

• Base model: DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1
• Runtime / artifact format: NInfer
• Calibration datasets: NVIDIA Open-SWE-Traces, Thoughtworks Agentic Coding Trajectories, Nebius SWE-agent trajectories