The Mash · three stories, one thread
Local AI workflows, training queues, and compatibility
Ollama adds LFM3 support on Linux and Windows, Unsloth adds training and browser controls, and Transformers fixes a configuration setter.
Explore
Test LFM3 text-to-speech in Ollama on Linux or Windows. Separately, check how Unsloth handles a cancelled job in a multi-LoRA queue, and test the Transformers setter fix against the affected vLLM loading path.[1][2][3][4]
Create
Make a test log for the three separate workflows this week. Record an Ollama LFM3 audio run, an Unsloth queued-job cancellation, and a Transformers configuration check. Do not assume these tools work together.[1][2][3][4]
Theorize
Hypothesis: builders may need to test workflow boundaries more often as model support, training controls, and configuration fixes ship separately. These claims do not establish interoperability between the tools.[1][2][3]
In the mash
Ollama adds LFM3 support on Linux and Windows
Ollama v0.40.13 lets Linux and Windows users run Liquid AI’s LFM3 models, including text-to-speech in lfm3-audio and lfm3-omni.
Unsloth adds multi-LoRA batching and browser controls
The beta release lets Unsloth users train compatible queued fine-tuning jobs together and let models control the built-in browser.
Hugging Face fixes rope_scaling setter in Transformers v5.21.1
The patch fixes a configuration setter that broke remote-code loading in vLLM.