AI developer tools · What shipped

Meta released Muse Spark 1.2 on August 5, 2026, its latest proprietary coding-focused AI model that scores 54 on the Artificial Analysis Intelligence Index, up from 51 for version 1.1. The update includes Muse Code, a terminal agent designed for repository-scale software work that can persist through crashes and coordinate tasks across up to 24 hours, with agentic knowledge-work performance jumping 260 Elo on GDPval-AA v2 benchmarks. Meta also released Muse Image and Muse Video in July, expanded to a one-million-token context window in 1.1, and priced the public API at $1.25 per million input tokens, positioning itself as a competitive near-frontier alternative to leading models at significantly lower cost.

Devin Desktop introduces SWE-1.6, described as the fastest coding model in the world, available with unlimited free access. The platform features an Agent Command Center where multiple agents handle coding tasks, a full-featured IDE with syntax highlighting, autocomplete, and debugging tools, and Fast Context for millisecond codebase searching. Key additions include the Agent Client Protocol enabling work across different models and agents, Spaces for sharing context between agents, and Supercomplete for predictive code completion. The desktop environment integrates with numerous tools via MCP Servers and extensions including Slack, Linear, ESLint, Prettier, Sentry, and others, positioning itself as the only IDE designed for seamless laptop-to-cloud handoff. The product serves over one million users and 4,000+ enterprise customers.

Amp shipped several developer productivity updates in late July and early August 2026. On August 7, improvements were released to reduce orb costs and efficiency. On August 6, the company introduced live reloading for testing agent changes in orbs. On August 4, file upload support was expanded to handle videos, logs, PDFs, and datasets. On July 29, Amp Labs partnered with Westpac for technology development. The platform also demonstrated model flexibility by swapping its default model without user complaints. Earlier on July 23, orbs gained event-driven capabilities to receive requests and react to external events.

oMLX 0.5.8.dev2 adds fixes for quadratic CacheList memory growth, head-of-line blocking in MTP requests, tool-call parsing issues in Ling and Inkling, and Codex launch isolation. The release also improves dual-stack networking, model compatibility, and localization support. oMLX 0.5.8.dev1 introduces Ling 3.0 Flash support with FP8 and mixed FP4/FP8 checkpoint support, thinking-budget enforcement during VLM MTP, and Jina Reranker v3.5 support. oMLX 0.5.7 completes DeepSeek V4 Flash 0731 prompt-format integration using the official reference encoder, improving tool schema handling, role transitions, and system context placement. The release fixes multiple issues including CacheList signature mismatches, MTP/PoolingCache boundary integrity, and long-context indexer fallback safety. oMLX 0.5.4 brings native DeepSeek V4 Flash 0731 with DSpark Lightning MTP achieving up to 85.6% faster code decoding, Inkling Small acceleration with 1.18–1.23x MTP speedups, and context benchmarking that measures usable context windows on the current Mac. Lightning MTP prompt priming improves Qwen3.6 decode by up to 18% with minimal prefill cost, and merged Gemma 4 checkpoints reach up to 1.97x faster decode with Lightning MTP.