Outlier Research
Collection
Ternary MoE overlays on Qwen2.5. 10B/40B/70B/150B scales. V3.3 active, V3.2/V2 archived. MMLU verified at n=14,042. • 8 items • Updated
⚠️ SUPERSEDED — use Outlier-Ai/Outlier-70B-V3.3 instead. These weights are retained live for reproducibility of earlier benchmark runs. All current research has moved to the successor.
Outlier 70B V3.2 (Superseded)
Earlier ternary MoE overlay on frozen Qwen2.5-32B-Instruct. Superseded by Outlier-70B V3.3 (alpha-fixed).
- Scale: 70B / ~32B active
- Base (frozen): Qwen/Qwen2.5-32B-Instruct
- Status: archival — do not use for new work
What changed
V3.2 MMLU was 81.49%. V3.3 alpha-fix adds +1.61pp from a 280-scalar, 15 KB overlay trained in 18 minutes on one B200 → 83.10%.
Historical benchmark (reference only)
| Benchmark | Score | Notes |
|---|---|---|
| MMLU 5-shot | 81.49% (n=14,042, lm-evaluation-harness v0.4.9.1) | Pre-V3.3 measurement |
Use the successor
- Current: Outlier-Ai/Outlier-70B-V3.3
- Research collection: Outlier Research
Why this is still public
ML research norms: earlier checkpoints stay live so external benchmarks and papers that cite this URL remain reproducible. This is not dead weight — it's the historical record.
Related
- Outlier desktop app: outlier.host — v1.4 shipping
- Discord: discord.gg/Hapennmdn9
Patents
Architecture covered by US provisional patents 64/026,886, 64/030,368, 64/034,028 (Kerr & Company LLC, 2026).
- Downloads last month
- 2,267