Qwen3 puts its reasoning trace and JSON tool call in one completion, open under Apache-2.0 on Fireworks
Synopsis
Open models previously forced a choice between showing a chain of thought and calling tools deterministically. Qwen3-235B-A22B places its reasoning inside a separate think tag and emits the exact JSON tool call in the same completion, so downstream code can ignore or audit the trace. The model is a 235-billion-parameter mixture-of-experts with about 22 billion active parameters, hosted serverlessly on Fireworks under Apache-2.0 and callable through an OpenAI-compatible client.
Interpretation
An open model can now return its reasoning and a tool call in a single request, with the reasoning block segregated so downstream code can ignore, audit or meter it. Until now, open-source LLMs forced a choice between showing the chain of thought or calling tools deterministically. The Fireworks-hosted Qwen3-235B-A22B returns the chain of thought plus the tool call on the first call, while the second call skips thinking and only makes the tool calls.
Thinking is switched by a call parameter rather than by swapping models or endpoints. Thinking mode is triggered when reasoning_effort is not none, generating a think-tagged reasoning block and a final answer, with recommended temperature about 0.6, top_p about 0.95 and top_k of 20. Non-thinking mode is triggered by reasoning_effort set to none or a /no_think tag, omits the reasoning block to save tokens and latency, and uses temperature about 0.7 and top_p about 0.8.
Scale is spread across a mixture-of-experts: 235 billion total parameters with only eight experts, about 22 billion active parameters, live per call. The model holds 128 experts and activates eight of them at inference. It runs serverlessly on Fireworks under Apache-2.0 and is reached with the official OpenAI client pointed at that endpoint, with everything else unchanged.
Perspective
The demonstrated scope is the Fireworks-hosted Qwen3-235B-A22B endpoint: switching between thinking and non-thinking modes, emitting the reasoning block and tool call in one request, and connecting through the official OpenAI client. It is aimed at engineers who need to control reasoning cost inside agent pipelines while keeping an auditable reasoning record.
The material gives no quantitative evaluation of the model on tool-calling or reasoning tasks, and its performance description stays qualitative at near-frontier quality. It recommends at least 4k output tokens for long answers and up to 32k for book-length generation, which itself indicates that output length must be managed by the caller.