Merging two high-quality 9B reasoning finetunes produced an intermediate ~18B model that fills a practical gap: stronger behavior than each 9B source in many tasks while still running on consumer GPUs. The author stacked all 64 layers (32+32) and applied a focused 1000-step QLoRA heal to smooth the layer boundary — a low-cost step that noticeably improves structured outputs like code and HTML.
Key Capabilities
- Reasoning and chain-of-thought: Benchmarked at 40/44 (90.9%) on a 44-test capability suite covering reasoning, tool calling, agent workflows, structured output, and multilingual handling — matching or outperforming some larger MoE variants in those tests. This indicates the merge preserves and combines complementary reasoning strengths from both sources.
- Code & frontend generation: After the heal fine-tune, structured code and long HTML/CSS/JS outputs became reliably well-formed (stress tests produced multi-file frontend samples, 14–24K char outputs, and 62/63 checks passed). Useful when you need locally generated production-style frontend snippets or algorithmic code.
- Efficient local inference: Exported as a Q4_K_M GGUF (~9.2GB) with ~66 tok/s throughput in tests and a context length up to 262,144 tokens; configured to run on GPUs in the 12–16GB class (e.g., RTX 3060/4070) with modest VRAM tuning.
Who It's For & Trade-offs
Great fit if you want a locally runnable model that balances strong reasoning and code-generation with modest hardware: developers experimenting with agentic workflows, offline code generation, multilingual chat, or research into merge methods. Look elsewhere if you need a production-grade, safety-vetted model or comprehensive, reproducible evaluation — this is an experimental frankenmerge with known limitations (residual code-formatting artifacts, occasional missing fences, and not fully stress-tested for safety). The model was healed with only 1000 QLoRA steps, so some edge-case formatting failures and hallucination risks remain.
Where It Fits
Practically, this model sits between the original 9B finetunes and much larger 27–35B models: it aims to capture complementary strengths from two specialized 9B finetunes while remaining practical for local inference. Use it for prototyping, creative coding workflows, and experiments on merge/heal methods; avoid it for high-assurance production services without further evaluation.
