Nanbeige4.3-33B thinks by default and ships two sampling profiles. For agents, use temperature 1.0, keep preserve_thinking True, and call tools in XML, because that is how its benchmarks were run.
Nanbeige4.3-33B is one of the newest open weight models worth testing for agent work: Apache 2.0 licensed, with a context window of up to 524,288 tokens. But its model card makes clear that it is really two models in one, and the defaults most teams copy from their last deployment will run it the wrong way.
Start with reasoning. The model generates reasoning by default, and the enable_thinking parameter set to False switches it to non thinking mode. That default is right for hard problems and wasteful for classification, extraction or short chat replies, where you pay in latency and tokens for thinking you did not need. Decide per route, not per deployment.
Next, the sampling settings split by job. For agentic and tool use tasks, the card recommends temperature 1.0, top_p 0.95 and up to 65,536 max new tokens. For reasoning and chat, it recommends temperature 0.6, top_p 0.95 and up to 262,144 max new tokens. A single global temperature, especially a low one carried over from another model, will quietly degrade one of those two workloads.
The setting most teams will miss is preserve_thinking. It controls whether reasoning from earlier assistant turns is kept in a multi turn conversation. The card recommends False for general chat and True for multi turn tool use and code agent workflows, and it states that the published benchmarks were run with preserve_thinking true. If your agent loop strips prior reasoning, as many chat style integrations do, you are not running the configuration the scores came from, and long tool chains are where that gap will show.
Finally, match the tool format and server flags. The card says tool_call_format xml gives the best tool calling results, with json supported only for compatibility. Its vLLM and SGLang launch commands use the qwen3 reasoning parser and the qwen3_coder tool call parser with tensor parallelism of 4. Get those wrong and tool calls arrive as unparsed text, which looks like a model failure but is a configuration one.
Try it today
- Route simple tasks with enable_thinking set to False; leave reasoning on for hard problems.
- For agent and tool use routes, set temperature 1.0, top_p 0.95 and max new tokens up to 65,536.
- For reasoning and chat routes, set temperature 0.6, top_p 0.95 and max new tokens up to 262,144.
- Set preserve_thinking to True in multi turn tool use and code agent loops, and False for general chat.
- Use tool_call_format xml, and serve with the qwen3 reasoning parser and qwen3_coder tool call parser.