What is a decision model?

It is a model like Jev or Clef-Flash that does no text generation at all. Given input data plus prepared options, it outputs the probability of each option directly, which removes text-generation wait time and runs with lower latency than a conventional VLM.

Why did Qwen3.5-9B crash?

It judged 'turn right' correctly before the curve, but its text output was converted straight into control, so it cut the wheel to maximum. With about 0.7 seconds until the next inference, it kept turning and hit the inner wall.

How was the test run?

On a desktop with an Intel Core i7-14700F, NVIDIA GeForce RTX 4070 12GB and 32GB RAM, using Godot for the simulator and llama.cpp with Q4_K_M GGUF models: Clef-Flash (ggml-org) and Qwen3.5-9B (unsloth).

Large Language ModelsAutonomous DrivingZenn AI/ML7h agoOct 20, 2026, 13:00 JST

Clef-Flash finishes lap at 242.6ms; Qwen3.5-9B retires

Send on LINE

3 Key Points

What happened
In a Godot self-driving simulator on an RTX 4070, Cloudflare's Clef-Flash completed the lap at 242.6ms average inference versus Qwen3.5-9B's 704.2ms, which retired after hitting the inner wall.
Why it matters
Both models are 9B and Clef-Flash is fine-tuned from Qwen3.5-9B, so the gap appears to come from its decision-model design outputting option probabilities instead of generated text, enabling smoother steering.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The test isolates one variable: model size. Clef-Flash is built on Qwen3.5-9B, so both are 9B models run on the same quantized Q4_K_M GGUF format through llama.cpp, with identical vehicle settings and front-camera images fed to the AI each step. What differs is the output mechanism.

The simulator results show two effects compounding. Clef-Flash's 242.6ms per inference let it correct the steering about four times a second, and because each option comes back as a probability, the wheel angle and throttle strength could be scaled smoothly. Qwen3.5-9B's 704.2ms meant a large steering input stayed applied until the next result arrived, pushing it into the inner wall; after impact the camera view went white and it drove straight on.

Mizobe's conclusion is that for tasks where conditions change constantly in real time, a decision model's advantage is not only speed of judgment but the ability to adjust the strength of an action smoothly. All of this comes from a single simulated lap, not from road testing.

FAQ

What is a decision model?: It is a model like Jev or Clef-Flash that does no text generation at all. Given input data plus prepared options, it outputs the probability of each option directly, which removes text-generation wait time and runs with lower latency than a conventional VLM.
Why did Qwen3.5-9B crash?: It judged 'turn right' correctly before the curve, but its text output was converted straight into control, so it cut the wheel to maximum. With about 0.7 seconds until the next inference, it kept turning and hit the inner wall.
How was the test run?: On a desktop with an Intel Core i7-14700F, NVIDIA GeForce RTX 4070 12GB and 32GB RAM, using Godot for the simulator and llama.cpp with Q4_K_M GGUF models: Clef-Flash (ggml-org) and Qwen3.5-9B (unsloth).

Zenn AI/MLRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.