Releases·Tool Release
Unsloth adds Qwen-Image-2.1 support, custom Agent Skills, and 2x faster reasoning streaming
Unsloth released an update adding Qwen-Image-2.1 support for image generation, custom Agent Skills compatible with Claude Code folders, and chat/project management improvements. The release also doubles reasoning-block rendering speed and improves Linux desktop distribution with ARM64 and ROCm support. For practitioners, it consolidates fine-tuning, chat, and local model serving workflows into one desktop tool with more reliable training resumes.
6 engagement1 sourceTue, Sep 22, 2026, 03:46 PM
AI Overview
- Unsloth released Qwen-Image-2.1 support for image generation with a dedicated guide
- Users can add custom Agent Skills, reused from Claude Code and .agents folders, and select them in chat with @
- Long reasoning blocks now stream 2x faster at 60 FPS versus 30 FPS previously
- Linux desktop gains in-app Debian updates, a native ARM64 Ubuntu 24.04+ .deb installer, and a new AMD ROCm Docker image with Unsloth Studio and JupyterLab
- Chats are draggable in the sidebar, and Projects can be edited and organized from a dedicated Projects page
- Training improves: Hugging Face datasets work in recipes, past runs resume with correct remaining time, and dataset splits/subsets are respected
Related
Releases·Tool Releasefeatured
Qwen-Image-2.1 runs locally on 12GB VRAM via Unsloth GGUFs; Dynamic FP8 fits 6GB with offloading
Unsloth released GGUF quants and Dynamic FP8 support for Qwen-Image-2.1, letting the 7B image model run locally on consumer GPUs — 12GB VRAM via GGUF or 6GB via FP8 offloading. The team claims on-par performance with Nano Banana 2.0, and follow-up support in Unsloth Desktop fixed FP8/GGUF issues while adding image gen and editing.
Unsloth released GGUF quants and Dynamic FP8 support for Qwen-Image-2.1, letting the 7B image model run locally on consumer GPUs — 12GB VRAM via GGUF or 6GB via FP8 offloading. The team claims on-par performance with Nano Banana 2.0, and follow-up support in Unsloth Desktop fixed FP8/GGUF issues while adding image gen and editing.
Updated Wed, Sep 23, 2026, 12:10 AM — FP8 and GGUF quants now run properly in Unsloth Desktop, with image gen and editing supported (Sept 23).
2.2k engagement2 sourcesxTue, Sep 22, 2026, 04:14 PM
Releases·Tool Release
Jiaojie Skill: open-source AI context handoff across windows, models, devices
Jiaojie (交接) is an open-source Agent Skill that preserves AI conversation context across windows, models, devices, and languages, so users can switch models or windows without losing memory or restarting work. It supports multiple runtimes including Claude Code, Codex, Cursor, OpenClaw, and Gemini CLI, and is distributed under the MIT license.
Jiaojie (交接) is an open-source Agent Skill that preserves AI conversation context across windows, models, devices, and languages, so users can switch models or windows without losing memory or restarting work. It supports multiple runtimes including Claude Code, Codex, Cursor, OpenClaw, and Gemini CLI, and is distributed under the MIT license.
38 engagement1 sourcegithubThu, Aug 20, 2026, 08:57 AM
Releases·Tool Release
TensorSharp adds Qwen-Image 2.1 + LoRA support for local GGUF inference
The maintainer of TensorSharp, an open-source inference engine, added support for running Qwen-Image 2.1 locally for text-to-image generation and image editing, including LoRA adapters. The update ships GGUF model configs plus accelerated adapters like Pruna 8-step and Viggle Turbo that cut sampling passes. It matters to practitioners running image models entirely on local hardware without vendor runtimes.
The maintainer of TensorSharp, an open-source inference engine, added support for running Qwen-Image 2.1 locally for text-to-image generation and image editing, including LoRA adapters. The update ships GGUF model configs plus accelerated adapters like Pruna 8-step and Viggle Turbo that cut sampling passes. It matters to practitioners running image models entirely on local hardware without vendor runtimes.
24 engagement1 sourceredditSat, Sep 26, 2026, 08:26 PM
Releases·Model Releasefeatured
Alibaba releases Qwen-Image-2.1: open-weight 7B image generation and editing model
Alibaba's Qwen team released Qwen-Image-2.1 as open weights, a unified 7B model for both image generation and editing. It natively generates and edits RGBA transparent images, supports up to 10 reference images, and targets panoramas, infographics, and virtual try-ons. Day-0 support arrived from vLLM and ComfyUI, plus a live Hugging Spaces demo, giving practitioners an immediately usable open alternative to closed-source image models.
Alibaba's Qwen team released Qwen-Image-2.1 as open weights, a unified 7B model for both image generation and editing. It natively generates and edits RGBA transparent images, supports up to 10 reference images, and targets panoramas, infographics, and virtual try-ons. Day-0 support arrived from vLLM and ComfyUI, plus a live Hugging Spaces demo, giving practitioners an immediately usable open alternative to closed-source image models.
6406.5k engagement10 sourcesreddit, hackernews, xSun, Sep 20, 2026, 01:12 PM
Social·Community
User tests Qwen3.8 27b 1-bit quant on 8GB VRAM, finds it amusing
A user with 8GB VRAM tested the Unsloth 1-bit quantized version of Qwen3.8 27b, finding the degraded output humorous. The post highlights the trade-offs of extreme quantization for running large models on limited hardware.
A user with 8GB VRAM tested the Unsloth 1-bit quantized version of Qwen3.8 27b, finding the degraded output humorous. The post highlights the trade-offs of extreme quantization for running large models on limited hardware.
1.7k engagement1 sourceredditThu, Aug 20, 2026, 06:08 PM