0.8B open-source HaloGuard 2.0 claims streaming-guard SOTA, beating models 5-10x its size
const_reborn · x · 2026-11-05
The xnavkum team open-sourced HaloGuard 2.0, a streaming output guard that intercepts unsafe content token-by-token as the model generates — closing the blind spot where input guards never see what the model actually says.
Results: at just 0.8B parameters, it beats Qwen3Guard-Stream at every size and outperforms full-reply guards 5-10x larger, claiming SOTA among open streaming guards. Paper and weights are released; the work was led by a small team in the Bittensor ecosystem.
Related event: 0.8B HaloGuard 2.0 Tops Open-Source Streaming Safety Guardrails(2 posts)→
More from Models
- Cognition launches SWE-2 Flash, a 276B-A12B coding model that beats all US open-weight rivals — carlesgelada · 2026-11-05
- Kimi K3.1 Pro (2.8T params, 104B active) lands on Ollama Cloud with zero data retention — ollama · 2026-11-05
- Linear CEO Karri Saarinen: time to retire the word "context" in AI — round · 2026-11-05
- Codex on the $20 plan runs dry after about 2.5-3 hours of nonstop work — tristanbob · 2026-11-05
- Morph launches Reflex: first decision model with cached pricing, 10-20x cheaper than Jev — GabGarrett · 2026-11-05
- Hamel Husain: Benchmarks are used like credit scores, so criticism is warranted — HamelHusain · 2026-11-05