Releases·Tool Release
Unsloth Desktop adds command palette, packed 4-bit LoRA training, faster Laya decisions
Unsloth ships a Desktop update adding a command palette, shareable GGUF run settings, and clearer errors. The release trains LoRA directly on packed NVFP4, INT4, and MXFP4 checkpoints, cutting peak memory roughly in half, and speeds Laya decision models up to 4.1x.
4 engagement1 sourceThu, Oct 1, 2026, 02:06 PM
AI Overview
- Cmd/Ctrl+P opens a command palette; GGUF run settings can be shared as a link or saved without loading.
- NVFP4/INT4/MXFP4 checkpoints train packed: Qwen3.8-27B-NVFP4 peaks at 40.2 GB vs 72.9 GB on one RTX PRO 6000, with 11% shorter steps.
- Kimi-K2.7-Code trains on 6 B200s with INT4 experts packed, where a 16-bit copy would need about 2 TB; gpt-oss-120b trains on one B200.
- Triton kernels dequantize layers on the fly; INT4 stays exact instead of re-quantizing to NF4, which raised Qwen3-8B perplexity 1.6%.
- Laya decisions run up to 4.1x faster via hosted Decision API providers like TypeSafe; a 3-question guardrail check takes 2.8 ms vs 11.6 ms on B200.
- Unsloth adds FastModel fine-tuning for T5, T5Gemma, BART and Marian, plus an `unsloth eval` command running lm-evaluation-harness tasks.
- Clearer errors explain failed loads with a View logs button; the app gains zoom, Continue response, and Fix with the model for HTML canvas errors.
- Image models keep INT8/FP8 when offloading: Z-Image takes 2.3 s vs 38.4 s on a 10 GiB-capped B200; Qwen-Image-2.1 renders up to 3.6x faster on Radeon 8060S.
Related
Releases·Tool Release
Unsloth ships large performance release: Vulkan on AMD, MLX speedups, PyTorch 2.11
Unsloth added local-run support for Qwen3.8-Flash-Next and GLM-5.3-Flash with GGUF weights on Hugging Face, then shipped follow-ups enabling multi-token prediction for up to 2x faster inference and a major performance/bugfix release. The latest release switches AMD to Vulkan by default for a 20% prefill/decoding boost over ROCm and includes 200+ bug fixes. Practitioners gain faster local GGUF inference, one-click Hermes support, and MLX fine-tuning on Apple Silicon.
Unsloth added local-run support for Qwen3.8-Flash-Next and GLM-5.3-Flash with GGUF weights on Hugging Face, then shipped follow-ups enabling multi-token prediction for up to 2x faster inference and a major performance/bugfix release. The latest release switches AMD to Vulkan by default for a 20% prefill/decoding boost over ROCm and includes 200+ bug fixes. Practitioners gain faster local GGUF inference, one-click Hermes support, and MLX fine-tuning on Apple Silicon.
Updated Wed, Sep 9, 2026, 03:18 PM — September 8 release defaults AMD to Vulkan for 20% perf boost, fixes AMD gibberish issues, and adds 200+ bug fixes
1601.8k engagement7 sourcesgithub, xThu, Aug 27, 2026, 01:09 PM
Releases·Tool Release
Unsloth releases Auto Compaction and LAN Remote Access in latest update
Unsloth merged 200+ PRs into its unslothai/unsloth repository, adding experimental Auto Compaction for extended chats, preview Remote & LAN Access without Cloudflare links, and faster chat performance. The update also supports custom llama.cpp builds and introduces Unsloth Dynamic v3.0.
Unsloth merged 200+ PRs into its unslothai/unsloth repository, adding experimental Auto Compaction for extended chats, preview Remote & LAN Access without Cloudflare links, and faster chat performance. The update also supports custom llama.cpp builds and introduces Unsloth Dynamic v3.0.
6 engagement2 sourcesgithubThu, Aug 20, 2026, 04:58 PM
Releases·Tool Releasefeatured
Unsloth releases Unsloth Desktop, an open-source local app for running and training AI models
Unsloth released Unsloth Desktop, the first desktop app to run and train AI models locally, available on Windows, macOS, and Linux. The open-source app supports MLX, GGUF, diffusion models, and connects Claude Code and Codex to local LLMs, with claims of 2x faster training and 70% less VRAM usage.
Unsloth released Unsloth Desktop, the first desktop app to run and train AI models locally, available on Windows, macOS, and Linux. The open-source app supports MLX, GGUF, diffusion models, and connects Claude Code and Codex to local LLMs, with claims of 2x faster training and 70% less VRAM usage.
Updated Thu, Aug 13, 2026, 03:21 PM — v0.1.702-beta update adds tool calling for external providers, tunable VRAM, and 10% faster inference.
1301.6k engagement5 sourcesreddit, hackernews, githubTue, Aug 11, 2026, 03:31 PM
Releases·Model Releasefeatured
Unsloth releases Qwen3.8-27B Dynamic v3 GGUFs with 10% higher accuracy
Unsloth released new Qwen3.8-27B GGUFs using Dynamic v3.0, claiming over 10% higher accuracy on benchmarks like Div-300 and KLD. They also introduced 1-bit quants retaining 77% accuracy, enabling local runs on 8GB RAM. The update follows earlier support for running and fine-tuning Qwen3.8-27B locally.
Unsloth released new Qwen3.8-27B GGUFs using Dynamic v3.0, claiming over 10% higher accuracy on benchmarks like Div-300 and KLD. They also introduced 1-bit quants retaining 77% accuracy, enabling local runs on 8GB RAM. The update follows earlier support for running and fine-tuning Qwen3.8-27B locally.
Updated Fri, Aug 21, 2026, 02:15 PM — MagicQuant updated with Unsloth Dynamic v3, reporting substantial boost over v2.
42011.8k engagement6 sourcesgithub, reddit, xWed, Aug 19, 2026, 12:00 AM
Releases·Tool Release
unslothai/unsloth ships Docker, multi-user accounts, AMD support, and training updates
Unsloth shipped a major update to its open-source fine-tuning toolkit: a refreshed Docker image adding AMD RDNA1/RDNA2 support via a new ROCm image, multi-user accounts with isolation, INT8/FP8 image diffusion inference (2x faster), and ARM64 Windows CUDA support. The release also brings GRPO training improvements with Qwen3.5 support, MLX video input, and OS-level tool sandboxing on preview Linux/macOS hosts. For practitioners, this broadens Unsloth from NVIDIA-only setups to AMD and ARM workflows and makes shared GPU hosts more usable for teams.
Unsloth shipped a major update to its open-source fine-tuning toolkit: a refreshed Docker image adding AMD RDNA1/RDNA2 support via a new ROCm image, multi-user accounts with isolation, INT8/FP8 image diffusion inference (2x faster), and ARM64 Windows CUDA support. The release also brings GRPO training improvements with Qwen3.5 support, MLX video input, and OS-level tool sandboxing on preview Linux/macOS hosts. For practitioners, this broadens Unsloth from NVIDIA-only setups to AMD and ARM workflows and makes shared GPU hosts more usable for teams.
793 engagement2 sourcesgithub, xThu, Sep 17, 2026, 02:37 PM