Together Launches Post-Training Service for Custom Models via RL and SFT
togethercompute · x · 2026-10-20
Together AI announced Together Post-Training, a service for building custom models on its platform, supporting both RL from environment rewards and SFT from supervision data.
- Users define the training loop in code (sample → reward → forwardbackward → optimstep → weightssync); Together co-optimizes the underlying systems for model quality and training speed.
- Architecture: frozen shared base weights plus per-session adapter weights and optimizer state; outputs can be saved as checkpoints or deployed to dedicated inference endpoints.
- The illustrated workflow shows up to 5 experiments on one model resource, evaluation through a Multi-LoRA endpoint, then merging and quantizing the winning adapter (e.g. to FP4) for production serving.
- Positioning: in 2026 companies increasingly train custom models to cut costs and improve task-specific performance, without carrying the infra burden themselves.
Related event: Together Launches Post-Training Service for Custom Models via SFT and RL(3 posts)→
More from Infra
- REAX inference engine runs Perplexity Decider up to 6x faster, free for individuals — airesearch12 · 2026-10-20
- REAX engine to be continuously improved via open engineering competition — airesearch12 · 2026-10-20
- BitNet CPU inference is silently wrong: perplexity 6x reference, one-line fix works — matheus-c-feitosa · 2026-10-20
- Data centers are the real paperclips and the global bond market is the steel — m2saxon · 2026-10-20
- AMD: World now consumes 40 quadrillion inference tokens monthly, up 95x in two years — Beth_Kindig · 2026-10-20
- Meta details Prometheus gigawatt-scale supercluster, targeting ~1M GPUs — BenBajarin · 2026-10-20