Continual learning on Qwen3.5-397B is said to match Opus 4.8 for about $450k
josh_wills · x · 2026-07-28
Continual learning on Qwen3.5-397B is claimed to reach Opus 4.8-level performance for about $450k
The post highlights a new finding for SovereignAI: applying a π-shaped continual learning method to an open-weight model, specifically Qwen3.5-397B, reportedly yields a genuine frontier model competitive with Opus 4.8.
What the claim says
- The method is applied to an open-weight base model rather than a proprietary frontier model.
- The reported result is frontier-level competitiveness at roughly $450,000 in compute.
- A full technical report and open-source models are said to be coming soon.
Why it matters
- If validated, it suggests that open-weight models can be pushed much closer to frontier capability with a specialized continual-learning recipe.
- The post also signals a likely upcoming technical report plus open-source release, which would make the result easier to inspect and reproduce.
Related event: SovereignAI Boosts Qwen to Near-Opus for $450K(4 posts)→
More from Models
- Kimi K3 Held Top Spot for Less Than Two Weeks Before Being Overtaken in AI Model Rankings — Angaisb_ · 2026-07-28
- Claude Opus 5 Max takes No. 1 in Arena’s frontend coding leaderboard — scaling01 · 2026-07-28
- ProteinGym-LLM ranks Claude Opus 5 highest on a 217-task protein variant benchmark — LeoTZ03 · 2026-07-28
- Kimi K2 Third-Party Inference Pricing Matches Official API as Location Becomes Key — kevinsxu · 2026-07-28
- Arena’s new factuality ranking puts Claude Opus 5 Max at #1 — arena · 2026-07-28
- Kimi K3 used 51.2 million sandboxes across 1.5 million images — tarantulae · 2026-07-28