Summary: Together AI announced on October 26, 2026 that MiniMax M3.1 Flash is now available for post-training on its platform — you can run SFT or reinforcement learning with your own data, rewards, and environments while keeping the training loop in your own Python code. For builders shaping an open model for a specific agentic workload, this turns a cheap, long-context inference model into a trainable one without managing GPU infrastructure.
What Together AI announced
The announcement from Together AI is short on ceremony: take M3.1 Flash and shape it for the agentic workload you need. The pitch is that your own code owns the loop — batching, loss, learning rate — while Together AI runs the sharded trainer, sampling engines, and weight sync on remote GPUs.
The Together AI changelog fills in the specifics, dated October 23, 2026: LoRA training on B200-SXM GPUs, a 262,144-token context for both training and sampling, LoRA rank up to 32, and up to three adapter sessions per resource. The cookbook renders the model’s prompts with the MiniMax codec. Full-weight training is not supported for this model.
Why the training loop staying in your code matters
Together AI’s post-training service is deliberately different from its simpler job-based fine-tuning. Per the post-training overview docs, billing is per GPU-minute rather than per token, and you control every step: inspect per-step metrics, reshuffle data mid-run, swap in a custom loss function computed in PyTorch from the model’s log-probabilities, or stop on your own logic.
That control is what makes the RL path in the announcement practical. Your code scores the model’s outputs, so you can train on any task where you can check the answer — including multi-turn tool-use environments, where the model acts over several turns. For agentic workloads where the reward is “did the task actually get done,” that’s the difference between guessing at SFT data and training directly on outcomes.
The same mechanism keeps iteration cheap: you can attach several LoRA sessions to one model resource, so a hyperparameter sweep costs one base model’s worth of GPU time rather than one per adapter. Each adapter registers as a model once it saves a checkpoint, so you can compare candidates on a single deployment before promoting the winner.
Constraints to plan around
The supported models catalog sets the boundaries. M3.1 Flash is LoRA-only with a rank cap of 32 and three concurrent adapter sessions per resource — fine for sweeps, limiting if you want heavy per-task capacity. Training context and generator context are both 262,144 tokens, which is generous for a trainer but well below the 1,048,576-token context the serverless inference version of the model offers. Post-training also runs only on B200 GPUs for now, with B300 and GB300 support promised later.
One practical note from the docs: keep GRPO rollouts within the trainer’s limit where generator context exceeds training context — here they match, so that particular mismatch isn’t a concern.
How this fits an M3.1 Flash stack
If you’re already serving M3.1 Flash on serverless — where it’s priced at $0.15 input / $0.60 output per million tokens with NVFP4 quantization — this closes the loop: train a LoRA against the workload you actually run, then deploy the winning adapter to a dedicated endpoint. We covered the model’s inference-side economics in MiniMax M3.1 Flash: 1M Context at Budget Token Rates; post-training is the missing half for teams whose bottleneck is task fit rather than context or cost.
If your needs are plain SFT on well-formed JSONL data, the simpler job-based fine-tuning service is still the right call — less code, per-token billing. Post-training earns its complexity when you need custom objectives, RL with your own rewards, or step-level control. If that’s your situation, the sensible first move is Together AI’s own advice: get your loop working on a small base model, then point it at MiniMax-M3.1-Flash.
Sources
- Together AI on X: minimax m3.1 flash is now available for post-training on together ai run sft or
- Supported models - Together AI docs
- Overview - Together AI docs
- Changelog - Together AI docs
View the original post on X