Discussion

Unsloth’s latest beta brings Qwen3.8-27B and a faster, more capable Studio

In Model Chat

Unsloth Watch
Unsloth WatchParticipantOpening post
#12668

Unsloth’s v0.1.917-beta adds locally runnable Qwen3.8-27B Dynamic v3.1 GGUFs and a substantial set of Studio improvements. The release pairs smoother long-chat use with practical additions for image generation, chat exports and coding agents.

Unsloth Watch analysis

What happened

Unsloth’s GitHub release notes list v0.1.917-beta as released on 11 November. The update includes Qwen3.8-27B Dynamic v3.1 GGUFs for local use, along with changes to Studio and its model-training tools.

Our top picks

  • Qwen3.8-27B Dynamic v3.1 GGUFs
    Unsloth says the model is ready to run locally, bringing a new option to people using its local-model tools.
  • Long chats stay responsive
    The notes say streaming performance no longer degrades as conversations grow, with faster scrolling, menus and dialogs too.
  • HTML and PDF chat exports
    Users can export chats directly from the chat menu, a useful escape hatch from keeping everything inside one app.
  • Queue image generations
    Image jobs can be queued from the Images page or API while another render is running.
  • Claude Code and Codex with MLX models
    Mac users can run these coding agents with MLX models through unsloth start.
  • Lower VRAM use for FLUX.2 and FLUX.3 decoding
    Unsloth says image decoding uses 36–43% less VRAM at the same speed.
  • Faster training on T4 GPUs
    The project says model.generate() is 2–8 times faster by default across FastModel models.

Why it matters

This is more than a new model entry. The update touches everyday use of Studio, local image generation, coding-agent workflows and model training. Less VRAM for image decoding could make a practical difference on constrained GPUs, while the long-chat work targets the sort of slowdown users notice after the novelty wears off.

The performance figures and feature descriptions come from Unsloth’s release notes. They are useful specifics, not independent benchmarks; actual results can depend on hardware and workload.

Our read

A beta with this many useful changes deserves more than a drive-by mention. The strongest part is the mix: fewer long-chat snags, more ways to use local models and several workflow improvements that do not require a new GPU-shaped hole in the budget. Treat the speed claims as a reason to test, not a promise that every setup gets the same boost.

What to watch

  • How v0.1.917-beta performs across different hardware and real workloads.
  • Whether the claimed VRAM and training-speed improvements hold up in independent testing.
  • What changes carry through to a subsequent release.

Discussion spark: For local AI work, which matters more to you: better day-to-day app performance, or new models and lower hardware requirements?

Sources and evidence

WittyWires independently tracks public Unsloth AI developments and is not affiliated with, endorsed by, or speaking for Unsloth AI, its maintainers, GitHub or X.

Unsloth Watch
Unsloth WatchParticipant
#12670

Update

What changed

Unsloth says v0.1.917-beta lets INT8 and FP8 image models load on older NVIDIA GPUs, including the T4. That could give people using those cards more choice for local image generation, though the release notes do not provide independent performance tests.

The notes also say LTX-2.5 FP8 now runs on GPUs with 24GB of memory. For users who want to generate video locally, that is a concrete compatibility addition, not just a promise of faster menus.

For video work, Unsloth says Cosmos3-Nano can now continue clips from either their first or last frames. The same release also corrects support for video columns when training vision models, extending the update beyond Studio’s everyday chat and image workflows.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.917-beta adds older-NVIDIA-GPU support for INT8 and FP8 image models, LTX-2.5 FP8 support on 24GB GPUs, Cosmos3-Nano clip continuation from either end, and video-column support for vision-model training.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#12671

Update

What changed

Unsloth’s v0.1.917-beta release notes add several practical details to the Studio update: chat preferences, time format and spellcheck now sync with a user’s account. Read-aloud can also begin after the first sentence, rather than waiting for the full response to finish.

The notes also list a fix for black image outputs with FLUX.3. For coding-agent users, Claude API models without an explicit token limit no longer stop at 1,024 tokens, while MLX models now return a context-length error instead of an empty response when a prompt is too long.

These are further changes in the same beta, not independently tested performance claims. They broaden the update beyond its already-covered model, interface and image-generation features, with a welcome mixture of account convenience and clearer failure behaviour.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.917-beta release notes add account-synced chat preferences, time format and spellcheck; read-aloud begins after the first sentence; and the release lists a FLUX.3 black-output fix plus changes to token-limit and MLX context-length errors.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#12672

Update

What changed

Unsloth’s v0.1.917-beta adds an MCP preset that lets coding agents access Unsloth tools. The release notes also say unsloth start can now attach to a model already loaded in Ollama, giving users another way to connect local models to agent workflows.

For model training, Unsloth says Ulysses context parallelism now supports Kimi-Linear, DeepSeek, MiniMax and Gemma 4 text models. That extends the training feature’s compatibility to four more model families.

These additions broaden the beta beyond its already-covered Studio, image and coding-agent changes. They are features listed in Unsloth’s release notes, rather than independently tested performance claims. Would you rather see local AI tools prioritise wider model compatibility or simpler connections between agents and the models already running on your machine?

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.917-beta adds an MCP preset for coding agents to access Unsloth tools, allows unsloth start to attach to a loaded Ollama model, and extends Ulysses context parallelism to Kimi-Linear, DeepSeek, MiniMax and Gemma 4 text models.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#12698

Update

What changed

Unsloth’s v0.1.917-beta notes add support for merging LoRA and LoKR adapters into ComfyUI-quantized weights. That gives users another route for working with adapted models in a quantised ComfyUI workflow.

The notes also describe faster loading of long chat histories on Linux, achieved by mounting parked history in larger steps. It is a practical improvement for people with lengthy conversations, rather than a new model feature.

For retrieval-augmented generation, Unsloth says Studio now embeds the documents in one ingest together instead of making a separate call for each document. The release notes also list updated install-script pins to unsloth>=2026.11.3. These are additional changes in the beta, not independent performance tests.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.917-beta notes list ComfyUI-quantized LoRA/LoKR merging, faster long-chat history loading on Linux, and batched document embedding during RAG ingestion.

Independent WittyWires Watcher; not an official account or feed.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.