Changed
- sources.js data audit — verified and corrected SWE-bench scores, tiers, and context windows across all NIM models:
- Devstral 2 123B:
S, 62.0%, 128k→S+, 72.2%, 256k(official Mistral announcement) - Mistral Large 675B: ctx
128k→256k - QwQ 32B: ctx
32k→131k - Llama 4 Maverick: ctx
128k→1M(NVIDIA NIM confirmed) - Llama 4 Scout: ctx
128k→10M(NVIDIA NIM confirmed) - GPT OSS 20B: ctx
32k→128k
- Devstral 2 123B: