Skip to main content
Flash trains a LoRA adapter on top of a curated catalog of base models. You set the base model with one line in your config, and list the live catalog from the CLI: 
flash models

model = "Qwen/Qwen3.5-4B"   # required — see `flash models` for the options
flash models is always the source of truth. The table below mirrors the current catalog.

​
Catalog

Every catalog model supports both SFT and GRPO, and all are hybrid reasoning (“thinking”) capable.
Model (model =)ParametersAlgorithmsReasoning
Qwen/Qwen3.5-0.8B0.9BSFT, GRPOHybrid
openbmb/MiniCPM5-1B1.2BSFT, GRPOHybrid
Qwen/Qwen3.5-2B2.3BSFT, GRPOHybrid
Qwen/Qwen3.5-4B4.7BSFT, GRPOHybrid
Qwen/Qwen3.5-9B9.7BSFT, GRPOHybrid
Qwen/Qwen3.6-35B-A3B35B total / ~3B activeSFT, GRPOHybrid
The default train.lora_rank = 32 is the safest choice across the catalog. If you raise the rank, validate before a long run: Flash checks that the resulting adapter can be deployed for the selected base model.

​
Serving prices

Serving is billed per token after deployment. Prompt and completion tokens have separate per-model rates, and cached prompt tokens use the model’s cached-input rate. The prices below are per 1M tokens; your Freesolo billing dashboard shows the charges you actually accrue.
ModelPrompt / 1MCompletion / 1MCached prompt / 1M
Qwen/Qwen3.5-0.8B$0.012$0.060$0.0024
openbmb/MiniCPM5-1B$0.012$0.060$0.0024
Qwen/Qwen3.5-2B$0.024$0.120$0.0048
Qwen/Qwen3.5-4B$0.036$0.180$0.0072
Qwen/Qwen3.5-9B$0.120$0.180$0.0240
Qwen/Qwen3.6-35B-A3B$0.180$1.200$0.0600
The Cached prompt column is the rate for prompt tokens served from the prefix cache, which is automatic. See Billing for how prefix caching works and when it applies.

​
Notes per model

  • Qwen/Qwen3.5-0.8B: the smallest Qwen3.5. Cheapest for smoke tests and fast iteration.
  • openbmb/MiniCPM5-1B: an on-device-class small model on a standard Llama architecture.
  • Qwen/Qwen3.5-2B: a small, capable step up when 0.8B underfits.
  • Qwen/Qwen3.5-4B: a balanced starting point for most tasks.
  • Qwen/Qwen3.5-9B: the largest dense catalog model, useful when quality matters more than the run cost.
  • Qwen/Qwen3.6-35B-A3B: a Mixture-of-Experts checkpoint (35B total, ~3B active per token) and the largest model in the catalog.
The Qwen3.5 entries are text-only fine-tunes: the checkpoints are natively multimodal, but Flash trains and serves the language model only.

​
Reasoning (“thinking”) models

Every catalog model is hybrid-reasoning: it can run with or without an explicit reasoning step. Turn reasoning on for a run with one line in your config: 
thinking = true
See the thinking field in the configuration reference.

​
Choosing a model

Start small

Use Qwen/Qwen3.5-0.8B or 2B to validate your setup and data cheaply. Get a run working before you scale.

Scale up for quality

Move to Qwen/Qwen3.5-4B or 9B once the task is wired up and you want stronger results. Larger models cost more per run.

Pick a model, then train

Set model in your config and submit a run.