Together launches Post-Training service: bring your own RL loop, they handle the GPU infra
togethercompute · x · 2026-11-13
Together AI launched Post-Training, a service where you own the training loop and Together handles the infrastructure. Customers like poetiq (co-CEO'd by former OpenAI board member Ian Fischer) use it to spin up complex RL research workloads quickly while keeping full algorithmic control.
- You control: recipe via API, data/prompts/environments, rewards/verifiers/judges, GRPO/PPO or custom losses, Multi-LoRA or full-weight training, deployment target
- Together handles: GPU capacity, training-rollout consistency for RL, parallelism setup, per-model rollout engines, async/sync support
- One RL step = sample() → reward → forwardbackward() → optimstep() → weightssync(); SFT drops the outer two
- Python-native loop, one-call GPU reservation, parallel LoRA sessions, Tinker-compatible, W&B tracking
More from Infra
- RL engineering debate: decoupled rollout vs colocated vLLM, staleness is not a rounding error — willcb · 2026-11-13
- Chutes' Parallax hits ~18,000 tok/s on one RTX 5090, ~1 cent per million tokens — bittingthembits · 2026-11-13
- LLM gateway latency benchmarked: 0.21ms median overhead but 89ms at p99 — jkriket · 2026-11-13
- Miles v0.2 Debuts at SGLang Summit: RL Training Across NVIDIA, AMD, TPU, Trainium — ying11231 · 2026-11-13
- A 14-minute video traces one ChatGPT question from the Enter key to the GPU and back — Glass-Garden-5888 · 2026-11-13
- Burkov's weekly: 94% of UK students use AI for assessed work, CUDA Rust, e-means — burkov · 2026-11-13