Rapid-MLX 0.11.0 brings 25.6x faster first-token latency to Apple Silicon agents
awnihannun · x · 2026-07-26
Rapid-MLX 0.11.0 adds a much faster local inference stack for Apple Silicon and positions it as something you can actually use for agent workflows.
- Prefix-cache reuse cuts a repeated 6k-token prompt from 13.1s to 0.51s TTFT, a 25.6× speedup.
- Response caching drops 656ms to 2ms with byte-identical output and no GPU decode.
- The project says it can hit 152 tok/s on Qwen3.5-4B-4bit, with cold starts in seconds.
- rapid-mlx chat now acts like a real agent: it can connect to any MCP server, chain tools autonomously, and show live tool activity.
- The release also adds support for five model families, including HY3 295B MoE, Qwen3-Coder-Next 80B, MiniCPM5, LFM2.x, and a 2-bit Ternary Bonsai model.
More from coding & agent
- A Claude workflow that loads context first makes the agent feel like a real second brain — evielync · 2026-07-26
- Teams are now running agents for weeks and months to finish tasks — ycombinator · 2026-07-26
- Talk says AI engineers should review code by risk level, not by ideology — altryne · 2026-07-26
- nslookup MCP Server: Empowering AI with DNS Lookups and Security Scans — modelcontextprotocol · 2026-07-26
- MCP server turns Indonesia’s data protection law into RAG-powered Q&A — modelcontextprotocol · 2026-07-26
- Developer Criticizes Codex Subagent Implementation: Memory Bloat and Crashes — mertdumenci · 2026-07-26