Skip to main content
Back

Qwopus3.8 Coder-Flash Brings a 27B Coding Agent to Your Local Machine

A 27B Qwen fine-tune distilled on 200K coding and agent trajectories, shipped as GGUF for local inference with llama.cpp, Ollama, and vLLM.

  • Qwopus3.8-27B-Coder-Flash is a coding-focused fine-tune of Qwen3.8-27B shipped as GGUF under Apache-2.0.
  • Trained on ~200K examples distilled from Qwen3.8-Flash-Max, Qwen3.8-Max, DeepSeek-V4-Flash, and NVIDIA Nemotron data.
  • Scores 93.87% on a local MMLU-Pro STEM subset using zero-shot chain-of-thought reasoning.
  • Passes 9/9 seeded coding-agent tasks and 135/144 strict-delivery episodes in error-recovery testing.
  • Decoder throughput of 8.97 tok/s, slightly behind the general Flash variant but ahead of the base model.
  • Available in Q4 through Q8 GGUFs from 16.8 GB to 29 GB for local llama.cpp, Ollama, and vLLM use.

Qwopus3.8-27B-Coder-Flash is a new coding-focused fine-tune built to read requests, call tools, edit files, run tests, recover from errors, and deliver finished artifacts. The checkpoint is available from its model page as GGUF quantizations under the Apache-2.0 license, with no approval required to download it.

GGUF packages quantized model weights for efficient local inference through runtimes such as llama.cpp, Ollama, LM Studio, and vLLM, although feature support varies among them. The release gives developers several memory and quality options for running a 27-billion-parameter coding model on local workstations.

The checkpoint fine-tunes unsloth/Qwen3.8-27B, which derives from Qwen/Qwen3.8-27B. Its author says the training recipe favors compact, directed reasoning over lengthy reasoning traces. The published benchmark results are author-reported and should be read as local evaluation data rather than independently verified scores.

The recipe draws from multiple teachers

Training used approximately 200,000 coding and complex-reasoning examples. Distillation converted outputs from larger teacher models into training data for the smaller checkpoint, covering three main areas:

  • Agent and tool use: About 20,000 examples, primarily distilled from Qwen3.8-Flash-Max, cover multi-turn tool calls, environment-state tracking, error recovery, and final delivery.
  • Algorithms and code generation: Examples from Qwen3.8-Flash-Max, Qwen3.8-Max, and DeepSeek-V4-Flash emphasize edge cases, refactoring, debugging, and implementation tasks.
  • Logical reasoning: Data from NVIDIA’s Nemotron series supplies multi-step reasoning and structured instruction-following examples.