Summary
According to a post on X by @OpenAgentsInc, two hours of tuning OpenAgents' native Coax + Metal 4 runtime on an M5 Max cut prefill for a 2.6k-token prompt on Inkling 2 Mini (16.4 GB, IQ4_NL) from 4m 4s to 5.0s. The post reports 522 tokens per second, 25% faster than llama.cpp. Long-context decoding rose from 10.6 to 35.7 tokens per second, and warm startup is 0.23 seconds. These are the company's own benchmark claims as described in the post.
AI-assisted summary based on the listed source.
What happened
OpenAgents says two hours of tuning its native Coax + Metal 4 runtime on an M5 Max (Inkling 2 Mini, 16.4 GB IQ4_NL) cut a 2.6k-token prefill from 4m 4s to 5.0s (522 tok/s, +25% vs llama.cpp). Long-context decode rose from 10.6 to 35.7 tok/s, and warm startup is 0.23s. Why it matters: Faster local runtimes make...
Why it matters
The post links the speed gains to Thinking Machines' new small model, saying faster local runtimes make it more practical on Mac laptops. Developers running local models on Apple hardware are the main audience, if the results hold up. Watch for independent benchmarks or wider release details before treating the comparison as settled.
Signal Intelligence
Signal Strength 80%
Technical label WATCH
Public Interest 50
Category OPEN SOURCE
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 82
Practical Impact Score 8
Novelty Interest Score 94
Consequence Score 10
Curiosity Score 16
Shareability Score 64
Why this is here
VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).