Shipped · AI and LLM tooling

What unslothai/unsloth shipped

The public repository of Unsloth · github.com/unslothai/unsloth

Written by FoxPlug from the repository's own releases; not affiliated with Unsloth. Paste your own repository in the box below and FoxPlug writes the same weekly update for your product. This page is an automatic summary of the releases, pull requests and commits in github.com/unslothai/unsloth; Unsloth did not write it and does not use or endorse FoxPlug. Every line links to the change it describes.

Get a weekly update like this for your product, free

Or Use it as a GitHub Action

Follow unsloth's weekly shipped digest

Week of October 5, 2026

What shipped

Why it matters

This week widened where the project can actually run: vLLM on AMD, prebuilt CUDA 13 kernels for Windows and ARM, and 4-bit loads on Intel XPU go beyond the NVIDIA-only path. Training also got more dependable with GRPO under FSDP2, applied rope_scaling, packed rows on the transformers path and host-RAM embedding offload. Many Studio fixes target users with smaller machines, mixed quantizers and older dependency sets.

Changelog entry

Example posts FoxPlug drafted from these changes. Not written or posted by the project.

Post for X

This week: vLLM installs on AMD Linux and WSL, prebuilt CUDA 13 kernel wheels for Windows and ARM, GRPO under FSDP2, rope_scaling actually applied, packed SFT via varlen flash attention, and plenty of Studio fixes for Metal, MIOpen and mixed quants.

Post for LinkedIn

This week the project broadened the hardware and training paths you can use. Studio can install and run vLLM on AMD GPUs on Linux and through Windows Subsystem for Linux, and prebuilt CUDA 13 wheels for flash-attn, causal-conv1d and mamba-ssm are now built for win_amd64, linux_aarch64 and win_arm64 and installed on Windows and ARM. Training saw GRPO run under an FSDP2 config, an explicit rope_scaling applied to the model and to vLLM, packed rows on the transformers modeling path routed through variable-length flash attention, and offloaded embeddings kept in host RAM. Studio also gained a run of reliability fixes: Metal fit within GPU wired headroom, torch.compile for AMD image generation, SQLite WAL deadlock gating, MCP call re-runs, and Intel XPU 4-bit loads without a C compiler.

Week of September 28, 2026

What shipped

Why it matters

This week brings major Studio audio features (Clone, Convert, Separate, Music) alongside deep performance work: Block Swap unblocks larger model training, streaming denoiser improvements cut inference time by significant margins, and first-start optimizations make the tool noticeably faster to use. Hardware-specific fixes for ROCm and Windows AMD also expand device support.

Changelog entry

Example posts FoxPlug drafted from these changes. Not written or posted by the project.

Post for X

Studio's audio workflows expand this week: Clone, Convert, Separate, and Music pages now available. Training gets Block Swap to stream transformer blocks from RAM. Image generation accelerates with fused kernels and async prefetch.

Post for LinkedIn

Unsloth Studio ships five new audio pages this week—Clone, Convert, Separate, and Music modes—alongside major performance gains. Block Swap enables training past VRAM limits by streaming transformer blocks from host memory. Image generation speeds up 50% on first start through kernel prebuilds and quantization probes. Denoiser streaming removes host waits for large models on B200 and L4.

Week of September 21, 2026

What shipped

Why it matters

This week brought major fixes for TRL 0.29 compatibility, performance improvements on desktop and multi-GPU Studio instances, and new features for attachments and skills in the UI. The batched MLX serving and response_format grammar support on MLX expand what the project can handle without backend routing.

Changelog entry

Example posts FoxPlug drafted from these changes. Not written or posted by the project.

Post for X

This week: TRL 0.29+ trainer patching, MLX response_format grammar support, Skills creation in Desktop, batched MLX serving, GPU split display in Studio, and faster multi-GPU nvidia-smi caching.

Post for LinkedIn

This week's updates focus on expanding inference capabilities and improving user experience across Unsloth Studio and Desktop. Major work includes full TRL 0.29+ support for experimental trainers, MLX grammar-constrained decoding, batched reply serving on MLX, and Desktop Skills management. Studio gains GPU split visualization and dramatic performance improvements on multi-GPU systems through nvidia-smi caching. Chat attachments now display as formatted cards with better file format support.

Weeks with too little public activity are left out rather than filled in. Last updated 2026-10-12.

Is this your repo? Ask us to remove this page.