Running an Opus-level coding agent locally: MTPLX claims 126.5 TPS on Mac
julianharris · x · 2026-09-17
Julian Harris recommends MTPLX as the best way to run local AI coding agents, linking a deep-dive claiming a free, locally hosted Opus-level agent at nearly 2x Claude Opus speed. The referenced release, MTPLX V2.11.3, reports peak 126.5 TPS on a MacBook running a Qwen model, averaging 61 TPS at 100k context and 50 TPS at 200k context. After users complained about output quality vs LM Studio, the team ran thousands of sample comparisons, fixed 8 exactness bugs, and claims bit-exact outputs at any temperature. The author also vents about Anthropic's rate limits and censorship as a driver for going local.
More from coding & agent
- When the AI Agent Builds the Tool Instead of Doing the Task — Similar_Job_6080 · 2026-09-17
- huggingface_hub 1.32 lets UV scripts declare runtime images for HF Jobs — vanstriendaniel · 2026-09-17
- Open-sourced Qwen-1B-RLCD runs type-safe JSON inference 5x faster on-device — JiliJeanlouis · 2026-09-17
- Microsoft event to demo open-source agent evaluation and governance tools ASSERT and ACS — davemccollough · 2026-09-17
- Cybersecurity Compliance Startup Comp AI Raises $34M Series A Led by Roo Capital and Grand Ventures — JosephJacks_ · 2026-09-17
- ChatGPT Chrome Extension Brings Codex Into the Browser With Full Tab Context — TawohAwa · 2026-09-17