Models¶
Choose the checkpoint and device topology before following a deployment command. The rows describe supported paths, not interchangeable CLI presets.
| Path | Checkpoint | Devices and placement | Decode | HTTP tools |
|---|---|---|---|---|
| Qwen3-14B | Hugging Face style checkpoint | Single device; TP for offline runs; DP replicas for HTTP serving | Standard generation | No registered tool parser |
| DeepSeek V4 MTP | Converted Flash W8A8 | 8 devices, overlapped DP8/EP8/TP1 | Autoregressive or MTP speculative | Function tools without structural constraints |
| DeepSeek V4 DSpark | Converted Flash W8A8 | 8 devices, DP2/EP8/TP4; or 16 devices, DP4/EP16/TP4 | Greedy target-only K=0 or speculative K=7 | Function tools; structural constraints on K=7 |
All three paths support offline generation through --prompt and HTTP completions/chat completions. DSpark currently requires greedy sampling; prefix caching and request limits depend on the path and speculative mode. See each deployment guide for those limits, Offline Inference and Tool Calling for common workflows, and the CLI Reference for shared arguments.