A tiny (0.6B) agentic coder built to run fast — including CPU-only VPS — via the Q4_K_M GGUF, while speaking the deterministic Hermes/ChatML <tool_call> format used by the Tiny-Giant harness.

Files

File	Description
`SteraQwen3-0.6B-Q4_K_M.gguf`	Q4_K_M quantization (~0.4 GB) — `llama.cpp` / Ollama / LM Studio, CPU-friendly
`SteraQwen3-0.6B-f16.gguf`	f16 GGUF — re-quantize to any level without retraining
`raw_weights/`	Full bf16 safetensors HF checkpoint
`val_meta.jsonl`	Held-out validation set shipped with the model

Training

Base: Qwen/Qwen3-0.6B (Apache-2.0, Qwen3 arch, ChatML-native)
Method: full fine-tune (not LoRA), bf16 + gradient checkpointing
Data: ~30k Tiny-Giant agentic tool-use / debugging conversations
Epochs: 2 · LR: 1e-5 (cosine, 3% warmup) · Seq len: 4096
Loss: full-sequence

Prompt format

Trained with an explicit ChatML / Hermes renderer. Pin the ChatML template when serving (--chat-template chatml); do not rely on auto-detection. Tool calls:

<tool_call>
{"name": "<function-name>", "arguments": {...}}
</tool_call>

Inference (llama.cpp, CPU-friendly)

llama-cli -m SteraQwen3-0.6B-Q4_K_M.gguf --chat-template chatml

License

Apache-2.0, inherited from the Qwen3-0.6B base model.

strykes/SteraQwen3-0.6B

SteraQwen3-0.6B

Files

Training

Prompt format

Inference (llama.cpp, CPU-friendly)

License

strykes/
SteraQwen3-0.6B