Local AI Stack Update: Unsloth, ComfyUI, and Ollama

Three releases this week impact local workflows: Unsloth adds support for Qwen3.8 and DeepSeek-V4, ComfyUI fixes nested latent sampling, and Ollama improves image transcoding.

Three of the stack’s core tools shipped meaningful updates this week, and all three land on your machine, not a vendor’s cloud. Here is what changed, the dates it really shipped, and what to do about it.

Model Training and Inference

Unsloth v73188 now supports local training and inference for Qwen3.8, Kimi K3, MiniMax-H3, Gemma 4, DeepSeek-V4, and FLUX. This expands the range of models you can fine-tune on a single GPU. The repository remains a primary tool for local LLM and diffusion model work.

Stable Diffusion Fixes

ComfyUI v0.33.1 addresses a bug in KSamplerAdvanced when add_noise is disabled on nested latents. Version 0.32.0 also improved debugging for nested tensors. These updates stabilize workflows that rely on complex latent operations.

Ollama Compatibility

Ollama v0.32.14 introduces WebP image transcoding for llama-server. It also allows Qwen renderers to tolerate non-leading system messages. These changes improve compatibility with specific model architectures and image formats.

What to do this week

Update your ComfyUI installation to v0.33.1 to resolve the KSamplerAdvanced latent issue. Verify your Ollama version is at least v0.32.14 if you use WebP inputs or Qwen models with non-standard system message placement.