TRL v1.17 ships Mistral Large 4 support: LoRA fine-tune a 1T model on a single B300 node
SergioPaniego · x · 2026-10-30
TRL v1.17 is out with: Mistral Large 4 support (the 1T-parameter "Le Chonk"), score centering by default in AsyncGRPO, tree attention for linear attention models, faster rollouts & packing for AsyncGRPOTrainer, cross-tokenizer distillation, and AMD GPU support. The team says you can now train a tiny LoRA adapter on a 1T model with a single B300 node.
More from Models
- OpenAI reverses cuts, doubles Codex usage limits for all Pro tiers — MicahBerkley · 2026-10-30
- Dream raises $17M seed from a16z to build spatial-reasoning AI for AEC — a16z · 2026-10-30
- Investor bets speech-to-speech will be massive, praises Phonic's model — graceisford · 2026-10-30
- User Praises OpenAI for Doubling Pro Usage and Including API in Pro Sub — Angaisb_ · 2026-10-30
- OpenAI doubles usage limits for Pro tiers and Business Premium, effective immediately — Polymarket · 2026-10-30
- Gemini 4 Argon tops 4 robotics tasks on RLE-Bench, ranks #2 behind GPT-6 Astra — daibond_alpha · 2026-10-30