OPEN LINK ↗
// 1h agoMODEL RELEASE
Ornith-1.5-122B Hits 84.5% SWE-bench
Ornith-1.5-122B is a 122B mixture-of-experts model claiming 84.5% on SWE-bench Verified and 78.5% on BrowseComp while running locally on a single DGX. Its MIT-licensed weights and OpenAI-compatible tooling target high-end private coding and agent workloads.
// ANALYSIS
Ornith is pushing open models toward frontier-level coding performance without requiring an API dependency, though the headline numbers still need independent validation.
- –Strong coding and web-agent scores make it compelling for private software-engineering workloads.
- –MoE architecture should improve efficiency relative to similarly capable dense models.
- –MIT licensing, Ollama support, llama.cpp compatibility, and tool calling reduce deployment friction.
- –“Local” still means serious DGX-class hardware, not a typical laptop or consumer GPU.
- –Independent SWE-bench Pro, Terminal-Bench, and real-world agent evaluations will determine how durable the claims are.
// TAGS
ornith-1.5-122bllmopen-weightsopen-sourcemoeai-codingbenchmarkinference
DISCOVERED
1h ago
2026-10-20
PUBLISHED
1h ago
2026-10-20
RELEVANCE
9/ 10
AUTHOR
0x0SojalSec