⚡ DavidAU-Qwen3.8-27B-TWIN-ColdFusion-Astra-709 (NVFP4)

High-Performance Agentic Coding & Reasoning Model by KatterMobile (Kaden)

❤️ Please leave a LIKE / HEART on this Hugging Face repository!
If you find this model, the fine-tuning, or the NVFP4 quantization useful, hitting the heart button at the top of this page makes a huge difference and helps independent developers like me stay motivated to release more frontier open-source models for everyone.

---

☕ A Personal Note from KatterMobile

"Hello! I'm KatterMobile (real name Kaden). I spent dozens of hours training, fine-tuning, merging, and quantizing this model on a single NVIDIA GB10 Grace Blackwell DGX Spark workstation. I am an independent researcher and developer without corporate backing. If you find this release valuable or use it in your coding workflows, please consider supporting me on Ko-fi—literally anything helps cover compute electricity and hosting costs! Thank you so much for reading and supporting independent AI."

☕ Support on Ko-fi: https://ko-fi.com/kattermobile
📬 Support & Inquiries: support@kattermobile.com

---

🌟 Model Overview

DavidAU-Qwen3.8-27B ColdFusion Astra NVFP4 is an unaligned, frontier-grade agentic reasoning and software engineering model. It combines the raw intelligence and uncensored problem-solving capabilities of DavidAU's Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709 base with deep multi-turn agentic fine-tuning on the 108,148-sample Astra coding dataset.

Quantized using NVIDIA's official ModelOpt 0.46 into native NVFP4 (4-bit floating point), this model runs with extreme speed and minimal VRAM overhead on modern NVIDIA GPUs (Blackwell GB10 / RTX 50-series / Ada Lovelace / Hopper).

---

⚙️ Mandatory Thinking & Sampling Parameters

[!IMPORTANT] Preserve Reasoning Traces: To obtain full reasoning performance and prevent premature cutoff, ensure your client or UI preserves <think> ... </think> tokens and exposes them in reasoning_content.

---

🧠 Transparent Architecture: Exactly How I Trained It

[DavidAU-Qwen3.8-27B-TT-ColdFusion-709 Base]
                    │
                    ▼
[QLoRA SFT on Astra Dataset (108,148 Samples, 5.59 GB)]
(16,384 sequence length, r=64, α=128, all linear projections, 250 steps)
                    │
                    ▼
[In-Place Safetensors Merge]
(All 12 base shards fused with trained LoRA deltas)
                    │
                    ▼
[NVIDIA ModelOpt NVFP4 PTQ]
(Hessian & calibration using Astra sequences, 3 shards: 19.24 GB)
                    │
                    ▼
[vLLM Serving with 1M Context & 512k Output]
(Triton attention backend, fp8_per_token_head KV cache)

1. The ColdFusion Foundation

• Base Model: DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored-NM-DAU-NEO-MTP-GGUF.
• DavidAU's ColdFusion models are renowned for their uncensored reasoning capabilities, multi-stage merge heritage, and strong code synthesis.

2. Multi-Turn Agentic Fine-Tuning (16k QLoRA)

• Dataset: 108,148 deduplicated, high-signal agentic samples (5.59 GB) from the Astra engineering suite (sft_all_astra_full.jsonl). Contains multi-turn autonomous tool use, terminal debugging, complex systems architecture, and mathematical verification.
• Hyperparameters:
  • Framework: Unsloth Studio QLoRA
  • Sequence Length: 16,384 tokens
  • Rank / Alpha: $r=64$, $\alpha=128$ (dropout: 0.0)
  • Target Modules: All linear projections (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, linear attention projections)
  • Convergence: Loss converged from 1.50+ down to 0.19–0.50 across 250 steps with anti-ruin validation.

3. In-Place Safetensors Fusion

• All LoRA deltas were merged directly in-place across all 12 base safetensors shards, preserving 100% parameter fidelity without runtime adapter overhead.

4. NVIDIA ModelOpt NVFP4 Quantization

• Calibrated via official NVIDIA ModelOpt 0.46 PTQ (hf_ptq.py) using real Astra calibration sequences.
• Packaged into 3 safetensors shards totaling 19.24 GB (down from 52 GB BF16).
• Supported by vLLM's FlashInferCutlassNvFp4LinearKernel for native FP4 tensor core GEMM acceleration.

---

🚀 Serving with vLLM (1M Context & 512k Output)

vllm serve KatterMobile/DavidAU-Qwen3.8-27B-TWIN-ColdFusion-Astra-709-NVFP4 \
  --served-model-name DavidAU-Qwen3.8-27B-TWIN-ColdFusion-Astra-709-NVFP4 \
  --host 0.0.0.0 \
  --port 8000 \
  --max-model-len 1048576 \
  --gpu-memory-utilization 0.88 \
  --max-num-seqs 10 \
  --max-num-batched-tokens 16384 \
  --enable-prefix-caching \
  --max-cudagraph-capture-size 32 --speculative-config \x27{"method": "mtp", "num_speculative_tokens": 1}\x27 \
  --kv-cache-dtype fp8_per_token_head \
  --attention-backend TRITON_ATTN \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --default-chat-template-kwargs '{"enable_thinking": true, "reasoning_effort": "xhigh", "preserve_thinking": true}' \
  --hf-overrides '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' \
  --override-generation-config '{"temperature": 1.0, "top_p": 0.95, "top_k": 20, "min_p": 0.0, "presence_penalty": 0.0, "repetition_penalty": 1.0, "max_new_tokens": 524288}'

---

💻 Quick Start with Pi Code CLI (pi)

Add to ~/.pi/agent/models.json:

{
  "providers": {
    "vllm-home": {
      "baseUrl": "http://<YOUR_HOST_IP>:8000/v1",
      "api": "openai-completions",
      "compat": {
        "supportsDeveloperRole": false,
        "supportsReasoningEffort": true,
        "thinkingFormat": "qwen-chat-template",
        "chatTemplateKwargs": {
          "enable_thinking": true,
          "preserve_thinking": true
        }
      },
      "models": [
        {
          "id": "DavidAU-Qwen3.8-27B-TWIN-ColdFusion-Astra-709-NVFP4",
          "name": "DavidAU Qwen3.8 27B ColdFusion Astra NVFP4",
          "reasoning": true,
          "contextWindow": 1048576,
          "maxTokens": 524288,
          "samplingParams": {
            "temperature": 1.0,
            "top_p": 0.95,
            "top_k": 20
          }
        }
      ]
    }
  }
}

---

📬 Contact, Support & Feedback

• ☕ Ko-fi Donations: https://ko-fi.com/kattermobile
• 📧 Support & Bug Reports: support@kattermobile.com
• 💬 Community Discussions: Feel free to open an issue or discussion thread right here on Hugging Face!

---

📜 Credits & Licensing

• DavidAU: For the incredible ColdFusion base model series (DavidAU on Hugging Face).
• Qwen Team (Alibaba Cloud): For the frontier Qwen architecture.
• NVIDIA ModelOpt & FlashInfer: For state-of-the-art NVFP4 quantization and CUTLASS FP4 GEMM kernels.
• Unsloth: For ultra-efficient 16k context QLoRA fine-tuning.

Licensed under the Apache 2.0 License.