Together launches post-training platform aligning RL training with serving
togethercompute · x · 2026-11-02
Together AI unveiled Together Post-Training, a service that keeps training and generation aligned: you own the recipe via a Python API (GRPO/PPO, custom losses, multi-LoRA or full-weight), while Together handles GPU capacity, training-rollout consistency, parallelism, and rollout engines tuned per model — so the reward you design is what the model actually learns, carried through to FP4 serving. One RL step = sample(), yourreward(), forwardbackward(), optimstep(), weightssync(). Includes parallel experiment sessions, multi-LoRA eval, Tinker compatibility, and cookbook recipes.
More from coding & agent
- Vercel engineer open-sources Security Audit skill you trigger via @bot reply — brandon_galang · 2026-11-02
- Vercel Labs shipped 95 releases in October, debuts agent-desktop for computer use — evilrabbit_ · 2026-11-02
- Claude gets praised for hard work but researcher jokes: at least call it "our laptop" — lateinteraction · 2026-11-02
- Agent ran same task in 40s vs 261s — the bug was his own requirement checker, not the model — Stunning-Sherbet1853 · 2026-11-02
- /be-ambitious: A Claude Skill That Pushes the Model Beyond Safe, Small Outputs — dr_cintas · 2026-11-02
- Dev recreates 'subvocal' for AirPods Pro 3 with one prompt in 9 minutes, open-sourced — _AustinCalvert_ · 2026-11-02