Oops! Something went wrong
[next-mdx-remote-client] error compiling MDX: Expected a closing tag for `<tool_call>` (4:103-4:114) before the end of `paragraph` 2 | 3 | ## What Sets It Apart > 4 | - Agentic-first design: the model is post-trained and tuned for agentic coding — it emits well-formed <tool_call> blocks and a chain-of-thought `<think>` trace that serving stacks can parse into separate reasoning_content and tool_calls fields, simplifying integration with agent frameworks. This means fewer engineering heuristics to map model outputs into tool invocations. | ^ 5 | - Large-context + multi-format checkpoints: supports a 262,144-token context window and ships as dense (9B) and MoE (35B, 397B) checkpoints with bf16, FP8, and GGUF variants for different deployment targets; the 9B dense checkpoint fits a single 80GB GPU while MoE checkpoints require multi-GPU tensor-parallel serving. This enables scaling from single-GPU experiments to production multi-GPU agents. 6 | - Open serving recipes and compatibility: official recipes target vLLM and SGLang (examples show tensor-parallel-size=8 for a single 8×80GB node) and the model exposes an OpenAI-compatible /v1/chat/completions endpoint with tool calling; recommended runtimes include Transformers ≥5.8.1, vLLM ≥0.19.1, and SGLang ≥0.5.9. More information: https://mdxjs.com/docs/troubleshooting-mdx
