Qwen3.8-27B-AgentCoder-MLX-4bit

🧠 Qwen3.8-27B-AgentCoder-MLX-4bit

A Fine-Tuned Model for Enhanced Tool Calling, Code Generation, and Reasoning

• Author: Bruno Pistone
• Repository: br1-pist/Qwen3.8-27B-AgentCoder-MLX-4bit
• License: Apache 2.0
• Base model: br1-pist/Qwen3.8-27B-AgentCoder

---

Model Description

Qwen3.8-27B-AgentCoder-MLX-4bit is a Quantized version with MLX (Apple Silicon) of the fine-tuned version of the Qwen/Qwen3.8-27B model, optimized for:

• 🧮 Complex reasoning tasks
• 🧰 Tool calling
• 💻 Code generation

The model was developed through sequential fine-tuning, followed by a Direct Preference Optimization (DPO) post-training stage to improve alignment, coherence, and reasoning accuracy.

Highlights

• Post-trained with DPO using chosen/rejected pairs for better alignment
• Excellent balance between tool use, code generation, and reasoning

---

🚀 Direct Use

Qwen3.8-27B-AgentCoder-MLX-4bit can be used directly for:

• ✅ Tool calling in complex reasoning tasks
• ✅ Code generation for Python, JS, and other languages
• ✅ Multi-domain reasoning (math, logic, Q&A)

⚠️ Out-of-Scope Use

• ❌ Highly sensitive or confidential data
• ❌ Domains requiring expert-level specialization
• ❌ Tasks where full explainability is mandatory

---

🧠 Training Details

Training Procedure

Phase 1 — Post-Training: Reinforcement Learning from AI Feedback (RLAIF) with Group Relative Policy Optimization (GRPO)

After sequential fine-tuning and preference alignment, the model underwent a GRPO phase in which an LLM judge — not a static preference dataset — supplied the reward signal. For each prompt the policy samples a group of completions, the judge scores each one against a rubric, and the advantage is computed relative to the group mean, so no value network is required.

Reward model (RLAIF)

• Judge: moonshotai/Kimi-K3 via the Amazon Bedrock Converse API
• Rubric-scored 0–10, normalised to 0–1; rewards evidence-gathering, correct tool choice, root-cause reasoning and verification, and penalises speculation, wrong-tool calls and symptom patching
• Unscorable completions (content filter, malformed verdict) are excluded from the group baseline rather than scored zero

GRPO parameters

• Generations per prompt (group size): 8
• Beta (KL coefficient): 0.04
• Loss type: dapo
• Reward scaling: group (unit variance within each group)
• Sampling temperature: 1.0
• Max completion length: 640 tokens, truncated completions masked out of the loss

Training parameters

• Learning rate: 1e-5, cosine schedule with a 10% floor
• Warmup steps: 46
• Per-device batch size: 2 × 16 GPUs × gradient accumulation 4 = 128 completions (16 prompts) per step
• Epochs: 1 → 230 optimiser steps
• Max grad norm: 0.5
• LoRA: rank 16, alpha 32, dropout 0.1, applied to both the full-attention (q/k/v/o_proj) and linear-attention (in_proj_qkv, in_proj_z, out_proj) projections
• Precision: bf16, FSDP full-shard
• Hardware: 2 × ml.p5.48xlarge (16 × H100), 21h47m wall clock

GRPO Data

• 3,686 multi-turn, tool-call conversations (87% multi-turn; prompts left untruncated)
• Coding, general Q&A, data management, bash execution

Objective

• Improve clarity, correctness, and helpfulness
• Reduce hallucinations and verbosity

Outcome

• Mean judge reward rose from 0.615 (first 25 steps) to 0.689 (final 25), +12.0%
• Mean completion length and truncation rate both fell, i.e. the gain came from concision, not padding
• Policy KL from the reference model rose from ~2.6e-4 to ~8.0e-3, confirming the policy actually moved

---

🖥️ Technical Specifications

Model Architecture

• Model type: Causal language model
• Parameters: 4.0B
• Context length: ~264K tokens
• Thinking mode: Enabled

Compute Infrastructure

Hardware

• GPU: NVIDIA H100 (80 GB VRAM)
• System RAM: 2 TiB
• Memory per vCPU: 10.67 GiB

Software

• Python: 3.12
• Transformers: 5.3.0
• Libraries: bitsandbytes, safetensors, torch, trl, scikit-learn, tokenizers, psutil, py7zr

---

🧾 Citation

BibTeX

@article{qwen3.5-4b-thinking-2507-toolcode,
  title={Qwen3.8-27B-AgentCoder-MLX-4bit: A Quantized version for Apple Silicon of the Fine-Tuned Model for Enhanced Tool Calling, Code Generation, and Reasoning},
  author={Bruno Pistone},
  year={2025},
  journal={Hugging Face Model Hub}
}

APA

Bruno Pistone. (2026). Qwen3.8-27B-AgentCoder-MLX-4bit: A Quantized version for Apple Silicon of the Fine-Tuned Model for Enhanced Tool Calling, Code Generation, and Reasoning. Hugging Face Model Hub. https://huggingface.co/br1-pist/Qwen3.8-27B-AgentCoder-MLX-4bit

---

🧭 Recommendations

• Tool use accuracy depends on task complexity
• Code generation may occasionally produce minor syntax issues
• Reasoning strongest in structured, logical, and mathematical contexts
• Avoid using this model for confidential or safety-critical applications

---

<p align="center"> <b>🧠 Qwen3.8-27B-AgentCoder-MLX-4bit — created by Bruno Pistone</b><br> <i>Enhanced reasoning, tool calling, and code generation — refined with DPO alignment</i> </p>