- Six training models: Added pretrained
Qwen/Qwen3.5-9B-Base, with its own adapter identity. Both 9B models now use H100 for hosted FP8 serving with a 64-sequence cap. See Models. - Spot training: SFT/GRPO can opt into Spot with automatic recovery checkpoints. See Spot training.
- Custom warm-start: Initialize a fresh custom optimizer from an existing adapter, or resume custom optimizer/algorithm state from an exact step. See Custom algorithms.
- Native tool RL: The two 9B models support parsed function calls and tool results in multi-turn GRPO, including thinking. See native tool calling.
- Private artifacts: Self-hosted controllers use Azure Blob storage for Flash-owned artifacts; Hugging Face remains optional for upstream access and explicit exports. See Self-hosting.
- Reliability: Improved serving rollout continuity, authorization recovery, custom-session attachment, and overloaded API request cleanup.
- Custom training: All five catalog models now support text-only custom
algorithms on one automatically selected Modal GPU. Use explicit
target_tokensfor soft-target distillation and losses over multiple tokens per position. - SFT tools: Include per-row OpenAI-style
toolsso training prompts carry the same function declarations as serving requests. Image-bearing rows do not support this field. - Longer context: Qwen serving now supports 262,144 tokens. Custom training remains limited to 4,096 tokens; SFT and GRPO also need a GPU shape that fits. See Context limits.
- Hosted GPUs: The catalog uses L4, L40S, H100, and H200 serving tiers. See the model table for each model’s tier and the current token prices.
- Token probabilities: Request generated-token scores and up to 20
alternatives with
flash models chat --top-logprobs Kor the Python client. See Top-K log probabilities. - SFT reliability: Multi-image datasets avoid the previous pixel-column overflow before training. Missing outputs and targets lost to formatting or truncation now have distinct diagnostics.
- Recovery: Improved checkpoint-deployment recovery and run submission handling across control-plane restarts.
- Custom algorithms: Write a local Python loop with async or blocking GPU
calls, custom losses, and durable resume. Start with
flash custom init. See Custom algorithms. - Small models: Qwen3.5 0.8B and 4B support text-only SFT, GRPO, and hosted serving. See the model catalog and prices.
- Adapter downloads: Download exact checkpoints directly with your API key, without a Hugging Face token. See Download an adapter.
- GRPO rewards: Read rollout token probabilities and explicitly control truncation masking.
- Images and transcripts: Image-capable models allow up to 32 images across
a conversation, including multi-turn observations. SFT can use stored
prompt_messages. - Quotes: Every algorithm now uses an authenticated control-plane quote. SFT reports actual dataset counts separately; packaged dataset profiling no longer imposes a separate 32 MiB cap.
- Environments: Managed packages use immutable published revisions;
flash env push,pull, and managed environment IDs remain the user workflow.
- DeepSeek: Use
deepseek-ai/DeepSeek-V4.1-Flash-1017for text-only chat, with optional thinking and a 32K context. See Supported models. - Rate limits: Hosted inference defaults to 600 requests per minute per organization. See Rate limits and retries.
- Progress: Run status shows the current stage and elapsed timings. See Reading progress.
- Reliability: Improved training startup, checkpoint saving, and serving.
- Hosted serving supports thinking-enabled checkpoints and base-model requests.
- Reasoning tokens count toward completion usage at the normal output rate.
- Read
completion_tokens_details.reasoning_tokensfor the reasoning count.
- Use
RUN_ID/finalorRUN_ID/step-Nfor deployment, chat, evaluation, export, and warm-start. Bare run IDs are not accepted. - Undeploying a checkpoint leaves other checkpoints available.
flash models deploy --waitdefaults to 4200 seconds.- The training catalog includes Qwen3.5 9B, Qwen3.8 27B, and Qwen3.6 35B-A3B.
- Training and customer-owned serving use Modal.
- Added customer-owned serving commands:
flash serve deploy,status, andundeploy. - Warm-start supports every pairing of SFT and GRPO.
- Logs identify separate attempts in retried runs.
- GRPO accepts
group_sizevalues of 2, 4, or 8, with at most 512 completions per step. - SFT estimates identify the published environment used for the quote.
- Checkpoint failures and out-of-memory errors include clearer diagnostics.
- Unknown CLI flags suggest the closest supported flag.
- SFT cost estimates return in the same command without starting paid training.
- Added explicit image limits and algorithm-specific image support.
- Added run-scoped chat and JSON/JSONL run status output.
flash env evalsupports split and parameter overrides.- Removed
model_revisionfrom training configs.
- Runs can automatically select multiple GPUs; use
--gpusto set a ceiling. - Use
prompts_per_stepfor GRPO;batch_sizeis SFT-only. - Environment IDs use
namespace/project/name. - Declare scorer dependencies with
[environment] pip. - GRPO requires
max_stepsormax_examplesfor a quote. - Added
flash-clias an alternate command name.
- Login warns about non-local HTTP addresses.
- Added self-hosting support and configurable data and log settings.
- Training configs reject settings that do not apply to the selected algorithm.
- Added
flash train --gpus Nand configurablelora_alpha. - Training requires a supported catalog model.
- Evaluate deployed models with
flash env evaland held-out suites inevaluations.py. - Use
flash models deploy --waitin scripts; it exits nonzero on failure. - Environment testing supports split and parameter overrides.
- Dry-run checks your balance against the quote.
- Record chat completions and export them as training data with
flash traces export. - API responses return reasoning in
reasoning_contentand answers incontent.
- Training configs and environment commands require a project.
- Run commands are grouped under
flash runs; serving commands underflash models. - Manage deployments and switch projects in the dashboard.
- Create and list projects with
flash projects createandflash projects list.
- Improved generation and SFT training performance.
- Multi-turn GRPO supports per-turn or per-episode scoring.
Added Qwen3.6-27B at the time of this release. See the current
model catalog for today’s supported models.