Training catalog
Every catalog model supports SFT and GRPO. Image and reasoning support depend on the model:
The 0.8B and 4B paths are text-only for both training and serving. For the other
three models, see Image inputs. Use
thinking = true to enable reasoning on
models with optional thinking.
Custom training is supported only on
Qwen/Qwen3.5-9B, text-only, on one H100. Its context and rank limits are lower
than the SFT/GRPO limits below.
A deployed checkpoint preserves its trained thinking mode, including on metered
hosted serving with Flash 1.3.9 or later. Reasoning tokens are included in the
normal output-token charge. See
reasoning support.
The catalog is the whole set of trainable models. A model outside it is
rejected when the config is parsed, before any GPU is rented, and there is no
config key that opts into an uncatalogued id. The error lists the supported ids.
Serving prices
Serving is billed per token after deployment, at the per 1M token rates below. See Billing for cached-input discounts. These are final customer rates, with no additional provider markup. Later price changes do not affect previously accepted requests.
The training catalog contains these five models. Retired 2B and Qwen3.6 27B
identities are rejected; Qwen3.8 27B is a distinct model, not a rename of an old
adapter’s base.
DeepSeek base-model serving
deepseek-ai/DeepSeek-V4.1-Flash-1017 is available as a hosted, text-only base
model through the OpenAI-compatible API. It does not support Flash training or
LoRA adapters. It is separate from the training catalog in flash models list.
The serving context is 32,768 tokens, including the rendered input, requested
output, and 64 reserved tokens. The upstream million-token window is not exposed
by this serving configuration. Thinking is optional and defaults to off;
reasoning tokens count toward output usage.
Check the hosted
/v1/models endpoint for availability. See Chat with DeepSeek
for supported request fields and an example.
Context window and LoRA rank
Every training-catalog model supports a 32768-token context for SFT, GRPO, and hosted serving.flash train enforces the applicable context and rank caps:
a config whose max_context_tokens exceeds the cap, whose GRPO prompt plus
max_completion_tokens cannot fit, or whose lora_rank exceeds the max rank, is
rejected. Custom training is limited to 4096 tokens and rank 64.
Choosing a model
Start with 9B
Use
Qwen/Qwen3.5-9B to validate your environment and dataset with a short run.Compare on your task
Evaluate
Qwen/Qwen3.8-27B and Qwen/Qwen3.6-35B-A3B on held-out data.
Compare quality and the quote for your configuration before scaling up.Pick a model, then train
Set
model in your config and submit a run.