Qwen3.8-27B Code-Tools-Merged

Full merged checkpoint of the Qwen3.8-27B coding/reasoning/tool-calling fine-tune stack.

Stack (merged in order)

1. Qwen/Qwen3.8-27B (base)
2. Akahsizrr/qwen3.8-27b-lora-xhigh-code-tools-1k (LoRA, merged)
3. Akahsizrr/qwen3.8-27b-lora-kimi-k3-tools-code-instr (LoRA, merged)
4. Akahsizrr/qwen3.8-27b-lora-comp-v1 (LoRA, merged)

Benchmark

LiveCodeBench v6 (test6.jsonl, 175 problems, pass@1, n=1, temp 0.2, top_p 0.95, max_tokens 32768, xhigh reasoning effort, vLLM 0.28.0):

• 3-adapter stack: 76.0% (133/175) with adapters 1+2 only; 75.4% (132/175) with comp-v1 included.

Usage (vLLM)

from vllm import LLM, SamplingParams
llm = LLM(model="Akahsizrr/Qwen3.8-27B-Code-Tools-Merged", dtype="bfloat16",
          max_model_len=36864, gpu_memory_utilization=0.90, enable_prefix_caching=True)
sp = SamplingParams(n=1, max_tokens=32768, temperature=0.2, top_p=0.95)
out = llm.chat(messages, sp, chat_template_kwargs={"reasoning_effort": "xhigh"})

Reasoning is emitted between <tool_call>/ ; extract the final code block from the answer.