[AI] My Fine-Tuned Model Got Faster By Being Rebuilt: Getting A LoRA Onto A 32GB Mac Without Paying For It
Situation In The Last Post I Fine-Tuned Muse Glimmer 30B To Call Tools Better, Found It...
Tag archive
Situation In The Last Post I Fine-Tuned Muse Glimmer 30B To Call Tools Better, Found It...

Nanbeige4.3-33B on a Mac Studio with MLX: measured speed, memory, a fit table and why our benchmark never started. Plus Particle-1 vs DeepSeek V4.1 Flash.

Our 4-bit DWQ build of Qwen3.8-Flash-Next scored 0.6932 on one M5 Max Mac Studio. The OpenRouter cloud run has 0.6846, and that 0.0086 gap is not proven.
Small Next-Concept Modeling models trained in MLX reconstruct TinyStories text almost perfectly but write worse text than an autoregressive baseline, and less training-time corruption makes it worse.

You can run HazyResearch's released Next-Concept Modeling checkpoints on the Apple Silicon GPU...

MiMo V2.6 Flash local on two Mac Studios (RDMA, MLX) scored 0.6605 on our agentic coding benchmark vs 0.7909 in the cloud. Where it lost, and the bugs.
Apple has a Foundation Models command-line tool now. When I saw fm serve in its help output, I wanted...
The web is bloated with Jev posts. The thing is a few weeks old, and my feeds are already full of...

The best local image generation model on a Mac, measured: six open models blind-scored and timed on an M3 Ultra, plus which licenses allow commercial use.
A production-focused review of vllm-mlx continuous batching, concurrent throughput, prompt-length behavior, and unified-memory limits on Apple Silicon.
Hello, everyone. Point a camera at something and get a sentence back describing what is happening...
Situation Been Looking At New Models Lately, And Came Across Muse Glimmer, A 30B Model...