AI developer tools · What shipped

llama.cpp added support for the Kimi-K3 text model with hybrid KDA (linear) + MLA (full) attention architecture, including cross-layer residual attention, latent mixture-of-experts, and other advanced features. The implementation includes specialized handling for MXFP4-pack-quantized expert tensors with lossless repacking, verified against reference implementations with 6.7e-05 relative error on final-position logits. Support also includes K3-specific chat formatting with reasoning tags, typed tool calls, and message delimiters, tested end-to-end on the full 2.8T model.

SWE Multilingual, a multilingual software-engineering benchmark for real-world code issue resolution across multiple programming languages, was updated August 15, 2026 with evaluations of 33 AI models. Claude Opus 5 leads the public snapshot at 89.5%, followed by Claude Opus 4.8 (84.4%) and Composer 2.5 (79.8%). The benchmark falls in the Coding category with a 20% weight in BenchLM.ai's overall scoring system but is currently displayed as reference only and excluded from the scoring formula.

oMLX 0.6.0 released with experimental distributed serving across Macs using tensor or pipeline parallelism, reaching 28.6 tok/s across two Macs versus 16.1 tok/s on one. The release adds community intelligence benchmark publishing, expanded Qwen3.8 and Gemma 4 support, improved long-context memory management with linear instead of quadratic CacheList storage, and concurrent prefill responsiveness with 1.6x to 43x decode throughput improvements. New model support includes Ling 3.0 Flash, Meta Muse Glimmer 30B VLM with DFlash speculative decoding, and Jina Reranker v3.5. DeepSeek V4 prefill is 23-28% faster across context lengths with optimized kernels. The release also adds built-in web search and speech-to-text capabilities to admin chat, improved reasoning controls, and serves as an open-source ML framework update with significant performance and feature enhancements.

OpenViking 0.3.22, an open-source context database for AI agents, has released benchmark results on long-conversation user memory (LoCoMo) and multi-turn agent tasks (tau2-bench). The system achieved 80–83% accuracy on user memory tasks across three agent integrations, up from 24–57% with native memory systems, while reducing input tokens by 34.3–91.0% and query latency by 58.45–66.10%. On agent experience benchmarks, memory integration lifted task success by +6.87 percentage points in retail and +11.87 percentage points in airline scenarios. The project is available for installation via pip and includes integrations with Claude Code, Cursor, Codex, LangChain, and other AI developer tools.