v3.4.0 @magnitudedev/cli@0.2.8
low risk
AudarAI/Audar-ASR-V1Oct 12, 2026 → Oct 12, 20260 days between releases
What's New
Focuses on Gemma 4 performance, GPU stability, and zstd memory protection
New Features
- Speculative decoding improvements
- Gemma 4 models paired with assistant models for faster generation
- Fallback to CPU for unsupported GPUs
Improvements
- Streaming embedding tables to reduce memory usage for large models
- Faster Gemma decoding after first-load optimization
- Improved zstd request handling to prevent memory bloat
Bug Fixes
- Fixed 'sampling distribution is nonfinite' errors on M1 Max and M1 Ultra Macs
- Fixed 8GB NVIDIA and AMD GPUs unloading models during long prompts
- Fixed model loading errors on computers without supported GPUs
Upgrade Notes
Catalog updates for Gemma 4 models are required to utilize assistant models. Ensure GPU drivers are updated for M1 Max/Ultra stability. Models on unsupported GPUs now run on CPU instead of failing.
The Two Releases
v3.4.0
v3.4.0
Oct 12, 2026· 10 days ago
This release focuses on improving store compatibility by adding evidence for read-only modes and fixing issues related to WAL-mode headers and ignored errors.
Full release notes@magnitudedev/cli@0.2.8
@magnitudedev/cli@0.2.8
Oct 12, 2026· 9 days ago
Focuses on performance improvements for Gemma 4 models and memory optimization, alongside fixes for GPU stability on Apple Silicon and various model loading errors.
Full release notes