Hayabusa 9B is a text-only, full-weight fine-tune of Qwen/Qwen3.5-9B for software debugging and structured agent action selection.
This repository contains a full merged BF16 Transformers checkpoint, not a LoRA, QLoRA, or adapter. It can be loaded directly by Transformers-compatible runtimes or used as the source checkpoint for MLX conversion.
Current Checkpoint
hayabusa-9b-temporal-v2-step0-r2-c8-rl
This checkpoint follows Round 2 Cycle 8 of the temporal curriculum. Cycle 8 replayed the clean Round 0 anchor reference, then applied DPO and checkpointed RL on samples from the current policy.
Artifact completed: 2026-09-06 UTC
Cycle reference: Round 0 clean anchor
DPO rows: 55
DPO policy-branch tokens: 1,121,120
RL rows: 56
RL policy sequences: 160
RL training tokens: 2,221,619
RL optimizer steps: 160
Reward policy version: 11
Context format: agent-optimized-debug-v3
Training context ceiling: 32K tokens
Serving context ceiling: 64K tokens
Format: merged BF16 safetensors, 32 shards
Assistant thinking: enabled and included in the training/runtime contract
For reproducible deployment, load the immutable Hub revision hayabusa-9b-temporal-v2-step0-r2-c8-rl rather than relying on mutable main.
Intended Use
Hayabusa is specialized for structured debugger contexts containing source, logs, traces, test outcomes, compact progress history, and constrained actions. It is not intended as a general-purpose chat model.
The checkpoint is suitable for download and conversion by MLX tooling because all model weights, tokenizer files, chat template, generation configuration, and safetensors index are present at repository root.
