AI developer tools · What shipped

MiniMax M2.7 leads the VIBE-Pro benchmark with a score of 55.6%, a repo-level code generation benchmark that tests end-to-end project delivery across web, mobile, and simulation tasks. The benchmark was verified on August 10, 2026, with 21 confirmed model releases in the preceding 30 days, reflecting ongoing development in full-project code generation capabilities beyond single-file snippets.

LiveCodeBench v6, a named competitive programming benchmark release, has 16 AI models evaluated as of August 10, 2026. Sakana Fugu-Ultra leads the public snapshot at 93.2%, followed by Sakana Fugu at 92.9% and Kimi K2.6 at 89.6%. The benchmark measures contest-style code generation and carries 20% weight in BenchLM.ai's overall scoring system, though v6 is currently displayed as reference only and excluded from the scoring formula. LiveCodeBench refreshes quarterly and maintains a public benchmark set; the top-10 range spans 21.2 points, indicating meaningful separation between systems.

Meta released Muse Glimmer, a 30B open-weight dense model with 120K+ context window optimized for local agentic AI workflows on NVIDIA platforms. The model delivers 20K tokens/sec on a single GPU and uses dense architecture to provide reliable instruction following and long-context coherence for multi-step agent tasks, with privacy-by-design as it runs entirely on-device without model sharding or external endpoints. Deployment options include vLLM, SGLang, NVIDIA NIM containers, and support across NVIDIA GeForce RTX 5090, DGX Spark, DGX Station, and Jetson edge platforms, with fine-tuning available via NVIDIA NeMo AutoModel supporting SFT, LoRA, and reinforcement learning.