# Phonon-2 Benchmark Scores & Performance

> Phonon-2 has Fermion speech evaluation results. Voice evidence is display-only and does not produce a text-model score.

Canonical page: https://benchlm.ai/models/phonon-2

Last updated: September 30, 2026

## Model Details

| Property | Value |
|----------|-------|
| Creator | Fermion Research |
| Source Type | Open Weight |
| Reasoning Type | Non-Reasoning |
| Context Window | N/A |
| Official model card | [Fermion Research model documentation](https://huggingface.co/FermionResearch/Phonon-2) |
| Overall Score | Not scored (voice evidence only) |
| Overall Rank | Unranked |

## Family & Coverage

- Family: Phonon-2
- Variant: base
- Benchmarks covered: 0 of 486
- Related earlier model: [Phonon-1](/models/phonon-1)
- Coverage note: Speech and voice results appear below; no weighted text-model benchmark rows are stored.

## Speech and voice evidence

[Phonon-2 launch evaluation](/voice-benchmarks/phonon-2-launch-evaluation)

Fermion ran Phonon-2, Phonon-1, and Parakeet Redux with the Open ASR Leaderboard code on full test sets. The other accuracy rows are leaderboard figures quoted by Fermion, without independent re-verification here. Some model cards report different runs; those values are not mixed into this comparison. Voxtral's download size is an estimate at 16 bits. Throughput includes log-mel processing and excludes model load; batch-128 GPU results are separate from single-stream results. The Mac runtime comparison uses 20 dictations totaling 797 seconds at each runtime's defaults. Download size does not measure runtime memory. These results do not enter text-model or voice-agent rankings.

| Metric | Result |
| --- | --- |
| Seven-set mean WER | 5.21% |
| LibriSpeech clean WER | 1.72% |
| LibriSpeech other WER | 3.92% |
| AMI WER | 9.37% |
| Earnings-22 WER | 6.96% |
| GigaSpeech WER | 8.35% |
| SPGISpeech WER | 3.70% |
| VoxPopuli WER | 2.46% |
| Download | 164 MB |
| Phonon-2 (MLX) | 174.0x realtime on M5 MacBook Air |

Fermion-run; Fermion launch table retrieved September 30, 2026. Lower WER is better; throughput excludes model load.

- [Evaluation source](https://www.fermionresearch.com/research/phonon-2/)

Open English speech recognition weights under CC-BY-4.0, derived from NVIDIA Parakeet TDT 0.6B v3. The 164 MB download stores encoder weights at about 2.1 bits; engines expand or requantize them at load, so download size is not runtime memory use. Runs through MLX on Apple silicon, CPU engines on Linux and Windows, and CUDA on NVIDIA GPUs. The launch page leaves the release day pending; the release ledger records September 2026.

## Other Fermion Research Models

- [Phonon-1](/models/phonon-1) - Score: not computed