I Added Day-One Muse Glimmer Support to Apple MLX-LM
5/5 next words matched the reference 0.9965 similarity out of a possible 1.0 🔩the handful of things this model does […]
I Added Day-One Muse Glimmer Support to Apple MLX-LM Read Post »
Local AI & Open Models
I Took Down Six of My Own Benchmark Videos. Then I Built the Local Agent Leaderboard.
64 s Local easy-suite time vs 122 s cloud 40 Checks every judge must pass first $0 Per token on
I Took Down Six of My Own Benchmark Videos. Then I Built the Local Agent Leaderboard. Read Post »
NVIDIA Shipped a Model That Sees and Hears — It Just Didn’t Run on a Mac. So I Wrote the Missing Piece.
NVIDIA gave away a 30B model that sees, hears, and reasons. The catch: only its text half ran on Apple Silicon. I ported the vision and audio towers to MLX, verified them against NVIDIA’s own code, and open-sourced it.
Why I Quantize Open-Weight Models for Macs — And Why Your Law Firm Should Care
Frontier cloud LLMs log your prompts. For privileged content — NDAs, medical records, source protection — that’s a deal-breaker. Open weights on Apple Silicon are the only honest answer, and there’s real, sustained demand from firms that nobody is talking about.
Why I Quantize Open-Weight Models for Macs — And Why Your Law Firm Should Care Read Post »
Hermes-4-14B Abliterated, MLX 4-bit — Apple Silicon Just Got Another Real Model
I converted Babsie’s BF16 abliterated Hermes-4-14B to MLX 4-bit so Mac developers don’t have to wait. 7 GB on disk, runs on a 16 GB Mac, drop-in for any mlx-lm or claude-code-local workflow.
Hermes-4-14B Abliterated, MLX 4-bit — Apple Silicon Just Got Another Real Model Read Post »
HumanEval on a MacBook — 81.7% pass@1, Wi-Fi off
Qwen 3 Coder 30B (8-bit MLX) scored 81.7% pass@1 on HumanEval running on a single M5 Max MacBook with the Wi-Fi off. Real run, all 164 problems, 14 minutes wall-clock. The first published number for this variant on this hardware.
HumanEval on a MacBook — 81.7% pass@1, Wi-Fi off Read Post »