Models & pricing.

DeepSeek V4 Flash (0731 instruct) on the shared pool: a 160K-token window with tool calling and prefix caching, billed per token.

DeepSeek V4 Pro on the shared H200 pool: an 800K-token window with tool calling and prefix caching reported as cached_tokens, billed per token. The larger of the two DeepSeeks the fleet serves; deepseek-v4-flash is the cheaper, shorter-context one.

Qwen3 0.6B on the shared pool: a small, fast base model billed per token, and the cheapest one to fine-tune — train a LoRA adapter on it from the training tab and call the result per token with no endpoint of its own.

Qwen3 4B Instruct (2507) on the shared pool: the fine-tunable base — call it per token, or train it on your own dataset from the training tab.

Qwen3.6 27B on the shared MI325X pool: a 256K-token window on a single GPU, with tool calling and separate reasoning, billed per token. The largest base the fleet can fine-tune on one card — train a LoRA adapter from the training tab and call the result per token with no endpoint of its own.

Qwen3.8 27B on the shared MI325X pool: a 256K-token window on a single GPU, with tool calling and separate reasoning, billed per token. Fine-tunable on that one card — train a LoRA adapter from the training tab and call the result per token with no endpoint of its own.

GLM-5.2 on the shared MI325X pool: the largest model the fleet serves, with a 128K-token window, billed per token. Not fine-tunable as an adapter — vLLM's fused MoE LoRA path cannot leave a KV cache on this base, so a tuned GLM-5.2 is registered as its own merged model instead.

Kimi K3 on the shared MI325X pool: a genuine 1M-token context with separate reasoning output and tool calling, billed per token. Accepts images as OpenAI image_url parts, up to 16,384 tokens each. Accepts reasoning_effort of low, high or max — not medium.

Kimi K3 with prefix caching: a repeated prompt prefix is served ~41x faster and reported as cached_tokens, which is what an agent loop resends every turn. Trade-off: recall degrades on very long context — measured 100% at ~98K tokens but 67% at 1M, and the failure is a confident wrong answer rather than an error, so this is capped at 128K. Use kimi-k3 for long-context work where accuracy matters.