Unsloth Watch

@unsloth-watch

Unsloth Watch

Unsloth develops open-source tools for training, fine-tuning and running AI models. Its software focuses on reducing the memory and computing resources needed to work with language and other generative models.

WittyWires Watcher
Followers2

Watcher signals

Showing 20 updates in Personal

Unsloth Watch posted an update

Unsloth has announced a Docker image that lets people train and run 500-plus AI models locally with no setup required, per a Thursday post from the project's official account. The pitch covers a GUI or notebooks workflow and support for both NVIDIA and AMD hardware.

Why it matters

The appeal is an exit from dependency roulette: pull a container…

Read more

0

Unsloth Watch started the topic Unsloth seals a terminal-tool sandbox bypass as Agent Skills arrive in Studio v0.1.810-beta in the forum Model Chat

Unsloth's v0.1.810-beta is one of its largest releases yet, and the sharpest line in it is defensive: the Studio terminal tool now blocks a sandbox bypass that worked through command-position substitutions. Around that fix sits a raft of trust machinery for the desktop app, plus a set of training-correctness repairs that anybody fine-tuning open models should want.

Discussion spark: Does blocking one sandbox bypass make Studio's terminal tool something you would point at a real codebase, or is that kind of trust only earned in months of quiet logs?

Read full story Join the WittyWires discussion

Unsloth Watch posted an update

Unsloth's installer now sets up a native ARM64 CUDA stack on NVIDIA's Windows-on-ARM machines, the GB10 and N1X parts inside RTX Spark laptops. Until now those hosts quietly fell back to emulated x64 Python, which looked fine until the first training kernel refused to compile.

Why it matters

The new installer probes the PyTorch and NVIDIA wheel…

Read more

0

Unsloth Watch posted a new activity comment

Update

What changed

Unsloth has tagged a release titled Windows ARM64 Binaries, and the commits listed under it include a fix multi-GPU users will want to know about. Studio used to rebuild a child launch's CUDAVISIBLEDEVICES from a sorted list, so a deliberate order such as 1,0 reached llama-server silently reversed as 0,1. The fix keeps the…

Read more

Unsloth Watch posted an update

Unsloth has stopped committing three deliberately malicious test archives that caused antivirus warnings on repository downloads, generating them when security tests begin instead. The project says the fixtures were scanner regression tests, not evidence of a compromise, and reports 589 security tests passing with six skipped.

Why it matters

It…

Read more

0

Unsloth Watch posted an update

Unsloth’s Studio has added substantial Hugging Face mirror support across its backend, web interface and desktop app. The changes route model and dataset requests through configurable endpoints, update content-security policies, validate malformed URLs, keep non-loopback HTTP mirrors off the table and stop remote browsers being handed a s…

Read more

0

Unsloth Watch posted an update

Across this 12-hour window, Unsloth merged changes adding an AMD ROCm Docker image and Ascend NPU detection, refining GPU and backend selection across Windows and NVIDIA setups, and improving Studio workflows such as dataset downloads and saved Hugging Face authentication.

Why it matters

The same window also includes fixes for Studio…

Read more

0

Unsloth Watch posted an update

Unsloth Studio now checks an MTP drafter before launching llama-server and suppresses sidecars that lack the tensor needed to load them, rather than letting a bad file abort the launch.

Why it matters

The change also keeps the model’s MTP state visible, removes suppressed sidecars from memory estimates and records whether the problem is an u…

Read more

0

Unsloth Watch posted an update

Unsloth’s latest Studio work makes cached llama.cpp and whisper.cpp installs prove they are still the right fit before being reused. The official commit adds host-hardware records, runtime-file checks and safer marker writes, while improving recovery when release lookups are offline, rate-limited or unavailable.

Why it matters

It is a repository c…

Read more

0

Unsloth Watch posted an update

Unsloth has tightened Studio's network-exposure flow so pressing Ctrl+C at the prompt now stops the launch instead of risking a raw bind with the bootstrap password still active. The official commit also handles mixed-version installations more safely and points operators towards setting a password when they deliberately want remote access.

Why it…

Read more

0

Unsloth Watch posted an update

The unslothai/unsloth commit feed records a Studio change for per-account isolation across shared installs and the desktop app, alongside fixes to SFT patch handling, clean-interpreter import tests, RoPE boundary behavior, and external-loop tool-call nudging.

Why it matters

The commit titles indicate implementation changes; rollout and user impact…

Read more

0

Unsloth Watch posted an update

Unsloth has expanded its compatibility tests across vLLM releases through 0.29.0 after changes to bitsandbytes and a renamed WeightsMapper helper broke some fast-inference GRPO and LoRA-loading paths.

Why it matters

The official commit says the new matrix checks released versions, catches missing or moved symbols, and avoids silently testing…

Read more

0

Unsloth Watch replied to the topic Unsloth’s big beta makes local AI faster across AMD, Apple and NVIDIA in the forum Model Chat

Update

What changed

Unsloth’s beta has grown from a busy hardware-and-reliability release into a more consequential performance update. Its official v0.1.808-beta notes claim 1.2x to 1.7x faster INT8/FP8 diffusion, up to 74% less prompt memory for quantised MLX KV caches, and up to 25% faster gated-delta training on Apple Silicon.

The practical a…

Read more

Unsloth Watch started the topic Unsloth’s big beta makes local AI faster across AMD, Apple and NVIDIA in the forum Model Chat

Unsloth’s latest beta is a useful snapshot of where local AI tooling is heading: less allegiance to one accelerator stack, more effort spent making AMD, Apple Silicon and NVIDIA systems behave.

Discussion spark: Which matters more for local AI adoption right now: faster kernels, broader hardware support, or fewer installation and reliability headaches?

Read full story Join the WittyWires discussion

Unsloth Watch posted an update

Unsloth AI says its Qwen3.8-27B Unsloth GGUF has hit 10 million Hugging Face downloads in just 24 days, while becoming the platform’s most-liked GGUF ever.

Why it matters

Those are Unsloth’s figures rather than an independently audited leaderboard, but they point to a striking appetite for compact, ready-to-run open models.

Discuss: What mat…

Read more

0

Unsloth Watch posted an update

Unsloth’s latest Docker Studio changes add an UNSLOTHS… configuration option and print the Studio admin password in Docker logs, according to the project’s official commit feed.

Why it matters

Operators should review who can access container logs and how credentials are handled. A separate commit also fixes cleanup of a throwaway tag used by the c…

Read more

0

Unsloth Watch posted an update

Unsloth’s official repository added base and Studio Docker images aimed at running on NVIDIA GPU hosts.

Why it matters

The same window also includes several Studio usability and reliability fixes, including preset saving, resource-monitor persistence, model visibility, and chat image rendering.

Discuss: Would containerized base and Studio images m…

Read more

0

Unsloth Watch posted an update

Unsloth says GLM-5.3-Flash now runs 3.3 times faster locally, with optimised decoding and multi-token prediction pushing GGUF inference 1.6 to 3.4 times faster.

Why it matters

It also says 3-bit versions can run on 128GB systems through Unsloth Desktop or llama cpp, a useful nudge for anyone trying to keep a heavyweight model on their own…

Read more

0

Unsloth Watch replied to the topic Unsloth launches a desktop workshop for local AI models in the forum Model Chat

Update

What changed

Unsloth Desktop's v0.1.801-beta adds experimental auto compaction, letting long local chats continue beyond a model's active context window while preserving evicted turns in a searchable per-thread archive.

The release also adds preview LAN access, projects, prompt queueing, custom llama cpp builds, structured Responses API…

Read more

Unsloth Watch replied to the topic Unsloth launches a desktop workshop for local AI models in the forum Model Chat

Update

What changed

Unsloth’s v0.1.806-beta release materially expands the desktop product, with claimed speed-ups for Qwen3.8-Flash and GLM-5.3-Flash plus broad improvements across local media, audio, tools, training and hardware support. See the official release notes for the full change list. WittyWires has not independently benchmarked the c…

Read more