Gemma

Anthropic pulled Fable and Mythos from the market following US intervention, a frustrating move considering Fable was performing for me like Opus was six months ago, before the steady performance degradation across versions 4.6, 4.7, and 4.8. It really makes me wonder where LLM usage needs to go to ensure we aren’t tied to these frontier labs.

A future path seems realistic with the continuous improvements in capabilities of consumer hardware like the new Apple M-series paired with MLX, the inevitable arrival of Apple foundation models as announced during the WWDC2026, and the rapid growth of open-source coding harnesses, local agent platforms like OpenClaw and Hermes, and tools like Ollama and LM Studio. It is only a matter of time before we gain real independence.

Performance and thermals were within my low expectations and I can see the Macbook Air M5 getting quite hot and throttling. I’ve been tackling simple tasks locally on my MacBook Air M5 (10C, 32GB) using LM Studio with Gemma 4 26B A4B QAT and the Zed editor, and the experience was remarkably similar to using an Anthropic model from a year ago. Honestly, seeing a mere 12-month gap between open-source LLMs on consumer hardware and the frontier labs gives me plenty of hope that within the next 12-24 months, I’ll be running something comparable to Opus 4.8 or Fable 5 right on my desk. It’s surprisingly relieving to not think about token consumption and usage limits; I feel more empowered this way.