Unsloth ships a local trainer that turns an LLM into a decision model
On 7 Oct 2026 (13:21 BRT) Unsloth posted that you can train a decision model locally with their open repo. The guide and the FastDecisionModel / DecisionTrainer code in unslothai/unsloth are the source. This is a training method, not a new model in the catalog: Unsloth did not publish the fine-tuned weights.
What it does
The LLM reads the state once. A small head, which the guide describes as the same design as Cloudflare’s Clef, scores every question and option together. The model does not write text. You can start from a plain LLM (unsloth/Qwen3.5-4B, Llama 3.2 3B, Gemma 4 E4B) or keep fine-tuning Laya or Clef. predict() returns choice, noul and score answers with probabilities. UNSLOTH_SYSTEMONE_MODEL=... unsloth studio serves the merged folder, and requests that ask for laya, default or jev-latest are answered by your run.
A follow-up commit on 8 Oct (94a00e9) adds the same training path on Apple Silicon through MLX.
Numbers to read carefully
The post says Qwen3.5 0.8B went from 20.7% to 74.3% aggregate accuracy on three decision benchmarks, on 4 GB of VRAM, with a Clef head, LoRA rank 64 and one epoch, and that downstream accuracy went from 30–37% to 78%.
The guide’s table is the breakdown of that aggregate: typed-decisions 36% → 73%, BANKING77 7% → 74%, CLINC150 19% → 76% (those three starts average 20.7; the three ends average 74.3). Holdout accuracy is 78% for the 0.8B, 81% for Qwen3.5 2B, 79% for Llama 3.2 3B, 77% for Gemma 4 E4B, 77% for a further Laya fine-tune. The mix is 12 public sets plus LocalLLaMA/typed-decisions. UnverifiedClaim.
The same page then tells you to match “our results” with 2 epochs and LoRA rank 16, not the rank 64 and one epoch in the post. The two settings are not the same run. There is no Hugging Face repo of the trained checkpoints (checked 8 Oct).
Why it is news, not a catalog page
The head is trained, so a finished run would be a decision model. Until someone publishes weights, the thing that exists is the trainer. Same treatment as other open recipes that stop at code.
Update, 9 Oct: free Colab notebook
Unsloth posted a free Colab notebook that trains Qwen3.5-4B into a decision model on about 8 GB of VRAM, using LocalLLaMA/typed-decisions (state, questions, gold answers). It runs 60 steps (about 1.7 epochs): about 2 hours on a free T4, 12 minutes on an RTX PRO 6000. Accuracy starts near 30% with the new head and reaches 75–78% on the test split in Unsloth’s runs (UnverifiedClaim). Calibration on held-out rows then prints an ECE. Still no published checkpoint.
Community run (10 Oct)
Akshay Pachaar posted a training video showing a local Qwen3.5 0.8B Jev-style fine-tune with Unsloth: 60 steps in about 10 minutes on 4 GB of VRAM. The post reports accuracy rising from 37% to 65%. The post does not link a checkpoint or give enough evaluation detail to compare this result with Unsloth’s benchmark table above. Treat the figures as UnverifiedClaim; this is another run of the existing training method, not a new catalog model. The author’s Jev explainer was published on 18 Sep and is background, not a new launch.