Not recommended. This is an experimental build. Use magi-1.0 instead.
Website: gehirns.pages.dev
magi-alpha-2 (retired experimental release)
Retired, not recommended. This supervised fine-tune did not improve on the untrained base: LiveCodeBench 48/150 against 50/150 for base Qwen3.5-9B Q4_K_M under the same think-enabled protocol, a difference within run-to-run variation (five runs of the same base scored 42 to 50). It is kept online for transparency. For a small, efficient model, see magi-1.0: no training, with 96% of BF16's benchmark score in 30% of the GPU memory.
9B coder build: Qwen3.5-9B base + SFT teaching diet v2, crushed to Q4_K_M (5.24 GB). From the GEHIRNS density program — maximum capability per gigabyte. Alpha quality: experimental, kept for the record.
- LiveCodeBench: 48/150 (think-enabled, greedy single-pass, measured 2026-10-16)
- Base: Qwen3.5-9B (Unsloth GGUF) + LoRA diet v2 (613 rows), merged → GGUF → Q4_K_M
- Context: 32768 (see Modelfile)
- Lineage: ex-
seele9-d2; registry aliasmagi-alpha-2-q4km
Run (Ollama)
ollama create magi-alpha-2-q4km -f Modelfile
ollama run magi-alpha-2-q4km
Honesty section
48/150 (32%) is a dead heat with magi-alpha-1's 49/150: diet v2 moved nothing. The supervised fine-tuning line was stopped on 2026-10-18, after a postmortem found mismatches between the training and evaluation protocols; a third diet was trained but not released. This artifact records the plateau point.