CUDA 13.3, Windows, Mac update

[danielhanchen]

Mini Update for multi platform support

To update Unsloth or install a new Unsloth Studio, you must use:

macOS, Linux, WSL:

curl -fsSL https://unsloth.ai/install.sh | sh

Windows:

irm https://unsloth.ai/install.ps1 | iex

Warning

DO NOT USE unsloth studio update since packaging will not get the latest updates

Mac Updates

• Re-enabled llama.cpp prebuilt binaries for Apple Silicon (M1-M4) - Mac OS 14 / 15 / 26 (Tahoe)
• Apple Silicon Mac OS 13 (Ventura) is source build
• Intel (x86_64) for Mac OS 13.3 / 14 / 15 / 26 (Tahoe) uses llama.cpp prebuilt binaries
• Intel for Max 13.0 - 13.2 is source build

Windows Updates

• CUDA 13.3 llama.cpp prebuilt binaries now work for Windows
• For CUDA 13.2, CUDA 13.1 and below, Windows devices uses CUDA 12.4 fallback - we'll work on CUDA 13.1 binaries soon.

CUDA 13.3 Update

• CUDA 13.3 non Linux binaries work. We'll still use CUDA 13.1 for now
• CUDA 13.3 solves the CUDA 13.2 gibberish problem - see llama.cpp CUDA 13.2 gibberish for IQ3_S / IQ3_XXS, IQ2_M - SOLUTION use CUDA < 13.0 or Unsloth Studio #4849

Blackwell GPUs Update

• For now Blackwell will have delayed releases of llama.cpp prebuilt binaries sine CUDA 12.4 does not work - we are working to resolve this soon.

What's Changed

• Bump install.sh / install.ps1 pin to unsloth>=2026.5.8 by @danielhanchen in Bump install.sh / install.ps1 pin to unsloth>=2026.5.8 #5791
• Studio: expose --parallel / -np flag on unsloth studio run by @danielhanchen in Studio: expose --parallel / -np flag on unsloth studio run #5737
• tests: unblock three stale assertions broken on main (MLX CI + Backend CI) by @danielhanchen in tests: unblock three stale assertions broken on main (MLX CI + Backend CI) #5803
• ci: install unsloth_zoo from git main in notebooks-ci + studio-backend-ci by @danielhanchen in ci: install unsloth_zoo from git main in notebooks-ci + studio-backend-ci #5802
• tool mask support by @Datta0 in tool mask support #5682
• Studio: add frontend i18n support by @penumbrazz in Studio: add frontend i18n support #5765
• Studio: unblock install on Linux ARM64 + Windows ARM64 + Intel Mac by @danielhanchen in Studio: unblock install on Linux ARM64 + Windows ARM64 + Intel Mac #5790
• Harden Linux llama.cpp prebuilt reuse by @mmathew23 in Harden Linux llama.cpp prebuilt reuse #5796
• Studio: expose image size setting in training UI by @Dariton4000 in Studio: expose image size setting in training UI #5743
• Studio: add configurable CPU thread pool limit by @alkinun in Studio: add configurable CPU thread pool limit #5760
• Trim README Advanced launch options blurb by @danielhanchen in Trim README Advanced launch options blurb #5809
• [RL] make sync weights conditional by @Datta0 in [RL] make sync weights conditional #4925
• Studio: add Gemini provider with web_search, code_execution, prompt caching, and Nano Banana image generation by @danielhanchen in Studio: add Gemini provider with web_search, code_execution, prompt caching, and Nano Banana image generation #5720
• Studio: add remote MCP server support by @NilayYadav in Studio: add remote MCP server support #5750
• Clear MRoPE after generation for GRPO by @Datta0 in Clear MRoPE after generation for GRPO  #5683
• Detect CUDA UMD Version from newer nvidia-smi output (fixes Windows installer does not detect CUDA from newer nvidia-smi output using "CUDA UMD Version", falls back to cu126 #5812) by @danielhanchen in Detect CUDA UMD Version from newer nvidia-smi output (fixes #5812) #5817
• Studio: remove dark mode upload circle by @Imagineer99 in Studio: remove dark mode upload circle #5813
• fix: honor --ctx-size and other forwarded args from unsloth studio run in Studio's context-fit logic by @mvanhorn in fix: honor --ctx-size and other forwarded args from unsloth studio run in Studio's context-fit logic #5815
• Fix non-streaming GGUF chat completion usage by @alkinun in Fix non-streaming GGUF chat completion usage #5781
• Fix/studio colab proxy and iframe - Unsloth Studio not loading in Colab (iframe "refused to connect" and wrong URL) by @LeoBorcherding in Fix/studio colab proxy and iframe - Unsloth Studio not loading in Colab (iframe "refused to connect" and wrong URL) #5844
• Studio: keep web search/code pills off on model load if user disabled them by @rolandtannous in Studio: keep web search/code pills off on model load if user disabled them #5851
• studio/setup.sh: cope with fresh CUDA toolkits like 13.3 by @danielhanchen in studio/setup.sh: cope with fresh CUDA toolkits like 13.3 #5826
• fix/strix halo and windows AMD ROCm support by @LeoBorcherding in fix/strix halo and windows AMD ROCm support #5301
• Version downpairing for Radeon+ROCm PyTorch wheels by @jaeiclee in Version downpairing for Radeon+ROCm PyTorch wheels #5353
• feature/use lemonade-sdk llamacpp-rocm binaries by @LeoBorcherding in feature/use lemonade-sdk llamacpp-rocm binaries #5303
• studio: ROCm cleanups follow-up to fix/strix halo and windows AMD ROCm support #5301 by @danielhanchen in studio: ROCm cleanups follow-up to #5301 #5874
• fix: wrong code block language for nightly Windows install by @mervivian in fix: wrong code block language for nightly Windows install #5877
• fix: bump openssl, rand, rustls-webpki, tar in Studio src-tauri for security advisories by @orbisai0security in fix: bump openssl, rand, rustls-webpki, tar in Studio src-tauri for security advisories #5866
• Fix error when context-resizing by @afsuyadi in Fix error when context-resizing #5860
• Update Studio sidebar logout icon to logout-05 by @shimmyshimmer in Update Studio sidebar logout icon to logout-05 #5878
• studio: pick a macOS llama.cpp prebuilt that loads on the host OS by @danielhanchen in studio: pick a macOS llama.cpp prebuilt that loads on the host OS #5883
• studio: make NVIDIA prebuilt selection track CUDA version bumps (Windows + Linux) by @danielhanchen in studio: make NVIDIA prebuilt selection track CUDA version bumps (Windows + Linux) #5879
• Remove dead top-level bitsandbytes import from models/_utils.py by @Rahul007007 in Remove dead top-level bitsandbytes import from models/_utils.py #5875
• ci(security-audit): make package installs network-resilient by @danielhanchen in ci(security-audit): make package installs network-resilient #5853
• Studio: add stdio MCP server support by @oobabooga in Studio: add stdio MCP server support #5863
• studio/frontend: fix MCP dialog overflow on long URLs by @oobabooga in studio/frontend: fix MCP dialog overflow on long URLs #5864
• Studio: clearer error for diffusion GGUFs loaded as chat models by @danielhanchen in Studio: clearer error for diffusion GGUFs loaded as chat models #5857
• Studio: harden stdio MCP gating and fix transport edge cases by @danielhanchen in Studio: harden stdio MCP gating and fix transport edge cases #5892
• Studio: fix phantom scroll area below messages with many sources by @danielhanchen in Studio: fix phantom scroll area below messages with many sources #5822
• Studio: pin the last pre-macOS-26 llama.cpp prebuilt instead of walking back by @danielhanchen in Studio: pin the last pre-macOS-26 llama.cpp prebuilt instead of walking back #5896
• fix(install): detect x86_64 Python venv on Apple Silicon and rebuild as arm64 by @ramankrishna in fix(install): detect x86_64 Python venv on Apple Silicon and rebuild as arm64 #5187

New Contributors

• @penumbrazz made their first contribution in Studio: add frontend i18n support #5765
• @Dariton4000 made their first contribution in Studio: expose image size setting in training UI #5743
• @NilayYadav made their first contribution in Studio: add remote MCP server support #5750
• @mvanhorn made their first contribution in fix: honor --ctx-size and other forwarded args from unsloth studio run in Studio's context-fit logic #5815
• @jaeiclee made their first contribution in Version downpairing for Radeon+ROCm PyTorch wheels #5353
• @mervivian made their first contribution in fix: wrong code block language for nightly Windows install #5877
• @orbisai0security made their first contribution in fix: bump openssl, rand, rustls-webpki, tar in Studio src-tauri for security advisories #5866
• @afsuyadi made their first contribution in Fix error when context-resizing #5860
• @Rahul007007 made their first contribution in Remove dead top-level bitsandbytes import from models/_utils.py #5875
• @ramankrishna made their first contribution in fix(install): detect x86_64 Python venv on Apple Silicon and rebuild as arm64 #5187

Full Changelog: v0.1.42-beta...v0.1.43-beta

---

This discussion was created from the release CUDA 13.3, Windows, Mac update.

[rehan243]

So we actually dealt with this exact kind of multi-platform headache recently. For our setup, managing prebuilt binaries across Apple Silicon, Intel Macs, and Windows machines was a massive pain. The specific issue with CUDA versions (like the 13.2 gibberish you're referencing) hit us hard on our inference nodes — we had to roll back to CUDA 12.4 for stability, and even then it wasn’t perfect.

The decision to stick to llama.cpp prebuilt binaries for Apple Silicon is smart, but re-enabling them specifically for M1-M4 while leaving older Ventura versions as source builds feels like a compromise workaround. Curious — does the source build on Ventura have any measurable perf difference against the prebuilts? We saw about a 15% gap when comparing prebuilt binaries versus source builds on M1 CPUs running PyTorch with custom ops.

Also, the fallback to CUDA 12.4 on Windows for older versions makes sense (we did something similar), but what's your take on the timeline for stabilizing CUDA 13.1 across all platforms? I'd guess that’s a blocker for Blackwell GPU support too, especially since prebuilt binaries are delayed there. We had to drop back to CPU inference on some Blackwell setups last quarter — not ideal, but it worked. Would love to know how you're prioritizing this internally or if there are any specific hurdles you're hitting.