Changelog
New models, faster training and serving, and platform updates from the Freesolo team.
Serve models on serve.freesolo.co
Deployed models are reachable on Freesolo's own serving domain, and the Flash CLI points there by default.
Cost-optimized GPU selection for every model
Each model now serves on the cheapest GPU that fits it, with per-model LoRA rank and adapter limits.
Validate environments offline with flash env test
Run a fixed set of episodes against your environment.py locally, with no GPU, before you push.
Live per-step training metrics in flash log -f
Follow reward, rollout truncation, and per-step GRPO metrics as a run progresses.
Faster on-policy distillation
Generate-ahead rollout prefetch overlaps generation with teacher scoring to cut idle GPU time.
7-10x faster responses on 9B and 27B models
CUDA graphs are enabled for mid-size models, sharply reducing decode latency.
Multi-GPU training jobs
Set gpu.count in your job spec to provision and price multi-GPU pods.
Per-turn or per-episode credit for multi-turn GRPO
Choose how reward is assigned across the turns of a multi-turn rollout.
Named reward metrics per step
Your custom reward metrics now show up per step in the training log.
SFT training up to ~48% faster
Activation shaping, cached tokenization, and chunked loss cut memory use and unlock longer context.
Live sample completions on the run page
Watch per-step GRPO and OPD sample completions stream in on the run detail page.
Documentation moved to docs.freesolo.co
Old /docs links now redirect to the new documentation site.
Train on images (multimodal)
Multimodal SFT, single- and multi-turn GRPO, and on-policy distillation for vision models.
Multimodal (image) chat serving
Serve vision models with image inputs passed straight through to the model, with correct EXIF orientation.
Chat a specific training step
flash chat RUN_ID/step-N serves that step's exact deployed revision.
Reliable warm-start from a deployed adapter
Warm-starting from an already-deployed adapter now retries transient throttling instead of failing.
flash env push uploads your full environment tree
Nested files beyond the root and dataset/ folders are now included.
Fixed a render loop in the new training-job composer
Qwen3.6-27B is now available
Train and serve Qwen3.6-27B; 9B and 35B serving prices have been refreshed.
Live run monitoring
Live training metrics, heartbeat progress, and auto-refreshing worker logs on the run detail page.
Welcome emails and a $50 credit offer
New users get a welcome email and a $50 credit offer after their first Flash run.
Flash CLI now runs on Windows
Guarded a Unix-only import that previously crashed the CLI on Windows.
More reliable recovery from GPU provider outages
Interrupted runs now converge to resumed or terminal instead of stranding.
32k context on served and trained models
Per-model context was raised to 32k tokens, and the 9B model moved to L40S GPUs.
Adapter export fixes
Correct key normalization for vanilla PEFT adapters, and clearer Hugging Face Hub error messages.