AI developer tools · What shipped

Claude Code version 2.1.233 was released with improvements to GitLab merge request support, memory management for Bash tool commands on Linux, WebFetch caching configuration, and numerous bug fixes. Key fixes include resolving cloud session connectivity issues, improving MCP v2 subscription handling, fixing permission notification hooks, and addressing Windows path validation for NTLM credential protection. Previous versions (2.1.232, 2.1.231, 2.1.229, 2.1.228, 2.1.227, 2.1.226, 2.1.225, and 2.1.224) included major features like subagent forking, cross-session messaging, self-hosted runner environments, and archive plugin sources, along with security hardening and performance improvements.

Google DeepMind released Gemma 4, a family of multimodal open models with text and image input capabilities. The release includes five variants: E2B (2.3B effective parameters), E4B (4.5B effective), 12B, 26B Mixture-of-Experts (3.8B active parameters), and 31B dense models, with context windows ranging from 128K to 256K tokens. Key advancements include configurable reasoning modes, improved coding benchmarks (31B achieves 2150 Codeforces ELO and 80% on LiveCodeBench v6), native function-calling support, enhanced multimodal processing with variable image resolution, and native system prompt support. The models feature distinct architectural designs optimized for on-device deployment on edge devices, workstations, and cloud execution.

oMLX is a local LLM inference server optimized for macOS with continuous batching and tiered KV caching. The tool supports text LLMs, vision-language models, embedding models, and rerankers on Apple Silicon, with versions available via macOS app DMG, Homebrew (brew install omlx), and source installation requiring macOS 15.0+, Python 3.11–3.13. Key features include an admin dashboard for real-time monitoring and benchmarking, OpenAI and Anthropic API compatibility, multi-model serving with LRU eviction and manual load/unload controls, and experimental multi-Mac distributed inference via MLX pipeline ranks. The project includes native custom kernels for GLM-5.2, MiniMax M3, and Qwen3.5 families (with reported 30x speedup for GLM-5.2 prefill: 845 vs 29 tok/s on M3 Ultra), tool calling with structured output, and per-model configuration profiles exposing models as <name>:<profile> variants without extra memory overhead.

Z.ai announced GLM-5.3, a new model with ~750B parameters available initially in their coding plan and coming soon to their API and Hugging Face in open weights within two weeks. The model demonstrates significant benchmark improvements, surpassing Moonshot AI's Kimi K3 on many benchmarks and matching or exceeding Claude Fable 5 and GPT-5.6-Sol on some metrics, achieving frontier-level performance on agentic coding benchmarks despite having only a third of Kimi K3's parameters. Z.ai achieved this through extended post-training rather than base model changes, using more environments, diverse tasks, and increased compute on RL training, with the company noting GLM-5.3 is their most capable model for cybersecurity tasks including vulnerability discovery and exploit analysis, though they're implementing staged release with security partner evaluation before broader API availability and full weight publication.