Microsoft Unveils Decision-1, Built on Qwen3.5-9B: 83.5% Accuracy at 85 ms

Microsoft has unveiled Decision-1, a compact model for fast classification and request routing. It is built on Qwen3.5-9B. According to the company, the model scores 83.5% accuracy across 36 benchmarks at 85 ms latency. There are no independent checks yet.

Why decision models matter

A decision model does not write long texts or reason out loud. Its job is to make a quick call: which category a request belongs to, which tool should get the task, which model should handle the query, whether it needs to pass through a filter. In agentic systems, these forks appear at every step. If each one costs a call to a flagship LLM, the bill and the latency grow in proportion to the number of steps.

That is why models with a few billion parameters suit routers and dispatchers well. Nine billion parameters is right in that range: the model fits on a single GPU and responds in a fraction of a second. Microsoft is entering a segment where other solutions already compete, and it brings its own task-specific optimization.

The choice of base model is telling, too. The company fine-tuned Qwen3.5-9B, an open model from Alibaba's family, rather than using its own architecture. For a narrow applied task, that is a rational path: the value is created in the data, fine-tuning, and optimization layers, not in pretraining from scratch.

What we know about the numbers

• Base: Qwen3.5-9B.
• Purpose: fast classification and routing.
• Accuracy: 83.5% across 36 benchmarks.
• Latency: 85 ms.
• Data source: Microsoft's own tests.

The figures should be read with caveats. 85 ms fits within the budget of an interactive scenario where the decision is made before the main response, but latency depends heavily on hardware, input length, and batch size. The announcement does not disclose the measurement conditions. It also does not say which 36 benchmarks made up the set or which models Decision-1 was compared against.

The same goes for accuracy. 83.5% means that, on average, roughly one in six decisions is wrong. For sorting requests by topic, that may be acceptable. For routing, where an error sends a request to the wrong place or lets dangerous content through, you need data on specific error classes rather than a single averaged number.

What to check before deployment

If you are building an agentic system or a request router, treat Decision-1 as a candidate for a pilot, not as a ready-made solution. A sensible order of checks:

• Run the model on your own traffic. An average result across 36 benchmarks says almost nothing about your categories and the language of your requests.
• Measure latency on your own hardware, and look at p95 and p99 rather than the average.
• Calculate the cost of an error for each type of decision, and decide where automatic selection is acceptable and where a human needs to review.
• Add a fallback path: when the model's confidence is low, send the request to a stronger LLM.
• Compare against baselines: a general-purpose model of the same size and a classical classifier. Simple approaches often deliver comparable quality for less money.
• Clarify the license, access terms, and deployment method. The original announcement does not include these details.

The very appearance of Decision-1 shows that major players have started to take the "dispatcher" layer of AI systems seriously. Until the model has independent measurements, it is a claim on a niche, not a proven advantage.