▌ quarterly report
MODELINDEX Q3 2026
Models ranked
31
Hardware tiers
1
Installations
0
Top score
69.1
Executive summary
- Q3 2026: 31 models ranked across 1 hardware tier, measured by 0 active installations on 1 hardware configuration.
- Highest score: Qwen3-Coder-30B-A3B-Instruct (IQ1_M) at 69.1 on Apple M4 Max · 128GB+ · 40-core GPU.
- On the most-measured tier, Apple M4 Max · 128GB+ · 40-core GPU, Qwen3-Coder-30B-A3B-Instruct (IQ1_M) leads.
- Quantization sweet spots: FP16 on 128GB+.
- This is the first quarterly report; trends start with the next one.
- 32 of 34 reports received this quarter carried a verified software signature (94.1%).
Top models per hardware tier
Apple M4 Max · 128GB+ · 40-core GPU
| # | Model | Quant | Score | tok/s | Installs |
|---|---|---|---|---|---|
| 1 | Qwen3-Coder-30B-A3B-Instruct | IQ1_M | 69.1 | 81.9 | 0 |
| 2 | Gemma 4 12B It | FP16 | 50.0 | -- | 0 |
| 3 | NVIDIA-Nemotron-3.5-Lightning-30B-A3B | Q8_K_XL | 49.3 | 16.4 | 0 |
| 4 | Qwen-AgentWorld-35B-A3B | FP16 | 47.4 | 15.8 | 0 |
| 5 | NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16.i1 | Q4_0 | 42.8 | 9.0 | 0 |
| 6 | nemotron-3.5-lightning-30b-a3b | BF16 | 40.6 | 8.8 | 0 |
| 7 | Qwen3-Coder-30B-A3B-Instruct | Q8_K_XL | 39.0 | 8.3 | 0 |
| 8 | NVIDIA-Nemotron-3.5-Lightning-30B-A3B | IQ4_NL | 38.2 | 21.7 | 0 |
| 9 | NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16.i1 | Q6_K | 38.0 | 35.9 | 0 |
| 10 | Qwen3-Coder-Next | Q4_K_M | 35.2 | 12.9 | 0 |
Quantization sweet spots
| RAM tier | Quantization | Avg score | Avg tok/s |
|---|---|---|---|
| 128GB+ | FP16 | 48.7 | 15.8 |
Top model pairs
No model pair reached the index this quarter.
Trends
This is the first quarterly report; trends start with the next one.
Data provenance
- Active installations
- 0
- Hardware configurations
- 1
- Reports received this quarter
- 34
- Signature verification rate
- 94.1%
Methodology
- Every figure is computed from anonymous performance reports that RunMyAI Desktop installations send automatically (Terms 6.9). A report carries model, quantization and hardware-specification measurements only -- no content, no identity.
- Only reports signed with the RunMyAI software-attestation key are ranked; the signature proves a genuine build produced the bytes.
- Reports are grouped by hardware tier: processor model, RAM tier and accelerator. A model is ranked on a tier from the measurements of every installation that ran it there.
- Speed outliers are rejected per model, quantization and tier with a modified z-score around the median (cut-off 3.5) before averaging.
- The 0-100 score weights delivered accuracy, sustained speed, time to first token and user ratings, renormalized over what was measured and capped by how much of it was measured.
- Quantization sweet spot: the quantization with the highest average score among the ranked rows of a RAM tier.
- Trends compare each ranked row with the same model, quantization and tier in the previous quarterly report.
See how these models run on your own machine.
RunMyAI Desktop measures every model you run, locally, and shows where it ranks on hardware like yours.
0 installations · 1 hardware configurations · updated daily·Methodology·Verify a report·Privacy policy