---
library_name: gguf
tags:
- gguf
- qwen3
- elite-dangerous
- covas-next
- function-calling
base_model:
- Qwen/Qwen3-0.6B
language:
- en
license: apache-2.0
---

# qwen3-0.6b-cn-gguf

Quantized GGUF export for a COVAS:NEXT fine-tuning experiment based on `Qwen/Qwen3-0.6B`.

This repo currently publishes the official `checkpoint-700` release from the local MLX GRPO training run.

This is still an experimental release, but `checkpoint-700` is the selected public artifact because it gave the best practical balance and felt less overcooked than later checkpoints.

## Status

- Official file: `Qwen3-0.6B-grpo-ckpt700-q8_0.gguf`
- Source checkpoint: `700`
- Format: `Q8_0` GGUF
- Intended use: local COVAS:NEXT experimentation and evaluation

## Selection Rationale

- Later checkpoints achieved slightly higher peak benchmark scores in some cases, but `checkpoint-700` was selected as the release candidate because it appears less over-optimized and more balanced in practical behavior.
- In particular, `checkpoint-700` preserves very strong tool behavior while avoiding some of the later-run drift seen in higher-update checkpoints.

## Benchmark Snapshot

Judge-based native Qwen eval snapshot for `checkpoint-700`:

- Overall: `240/276 (87.0%)`
- Tool calling: `96/108 (88.9%)`
- Event reaction: `31/42 (73.8%)`
- Contextual QA: `47/60 (78.3%)`
- Tool result summarization: `66/66 (100.0%)`
- Tool micro:
  - call made: `108/108`
  - name correct: `102/108`
  - args correct: `102/108`
- Avg case score: `0.807`
- Strict all-6 pass cases: `38/46`

## Notes

- Base model: `Qwen/Qwen3-0.6B`
- Project: COVAS:NEXT training experiments for Elite Dangerous ship-assistant behavior
- Exported from the local MLX GRPO training pipeline and converted to GGUF `Q8_0` for llama.cpp-style runtimes.
- See the source project experiment log for checkpoint comparisons and validation notes.