Oops! Something went wrong
[next-mdx-remote-client] error compiling MDX: Expected a closing tag for `<tool_call>` (5:152-5:163) before the end of `paragraph` 3 | ## What Sets It Apart 4 | - Self-scaffolding RL: the training objective jointly optimizes solution rollouts and the scaffold (task plans, tool calls, error-recovery routines). This reduces dependence on human-crafted agent harnesses and produces agent traces the runtime can parse. > 5 | - Explicit reasoning + tool-call outputs: by default replies include a `<think>` ... `</think>` block (available as reasoning_content) and well-formed <tool_call> blocks that servers can surface as OpenAI-style tool_calls, enabling tight integration with agent frameworks. | ^ 6 | - Single-GPU friendliness and long context: the dense 9B checkpoint is ~19GB in bf16 and runs on a single 80GB GPU; the family supports very long contexts (reported up to 262,144 tokens) and GGUF builds for local runtimes. 7 | - OpenAI-compatible serving: official recipes target vLLM, SGLang and Hugging Face Transformers, and the model exposes an OpenAI-style chat completions endpoint for immediate use with existing agent toolchains. More information: https://mdxjs.com/docs/troubleshooting-mdx
