Zircon v2
Zircon v2 is a 0.6B-parameter decision model from Fahrenheit Research. It runs fully on-device on Apple silicon. You give it an email, a message or a pending task plus a set of options, and it returns a calibrated probability for each option in under 50 ms per decision.
Specifications
| Version | Zircon v2 (0.6B, MLX 8-bit) |
| Released | September 2026 |
| Parameters | 0.6 billion |
| Model size | about 634 MB |
| Runtime | Apple silicon via MLX, runs locally |
| Output | One probability per option |
| Decision types | Pick one, yes or no, score on a scale |
| Response time | 30 to 45 ms per decision (1) |
Email and message triage (2)
| Task | Accuracy |
|---|---|
| What to do: reply, escalate, archive, or mark as spam | 94% |
| Notify now or hold | 100% |
| Priority on a 4-level scale | 93% |
| Pending steps: execute, hold for a reference, defer, or escalate | 100% |
Typed-decisions benchmark (3)
| Metric | Zircon v2 |
|---|---|
| Accuracy | 80% |
| Brier score (lower is better) | 0.043 |
Arithmetic reasoning (4)
| Task | Accuracy |
|---|---|
| Pick the correct answer to a word problem | 88% |
| Check whether a proposed answer is right | 86% |
(1) MacBook Pro (Apple M5), 8-bit weights, median. Speed varies by hardware. (2) Fahrenheit Research internal testing, September 2026, on held-out emails not seen in training. (3) Fahrenheit Research internal testing, September 2026. 400 cases (2,000 decisions) from the public LocalLLaMA typed-decisions test set. (4) Fahrenheit Research internal testing, September 2026. 400 held-out grade-school arithmetic word problems per task.
Built on Qwen3-0.6B (Apache 2.0). By Fahrenheit Research.
- Downloads last month
- 18
8-bit