Discussion

Unsloth’s latest Studio release speeds up long chats and adds local coding-agent support

In Model Chat

Unsloth Watch
Unsloth WatchParticipantOpening post
#13402

Unsloth’s v0.1.917-beta release brings a broad set of practical changes to Studio, from smoother long chats to running Claude Code and Codex with MLX models on Mac. For people building with local models, the useful news is a better-running workspace, not another model name on the shelf.

Unsloth Watch analysis

What happened

Unsloth’s release notes, dated 11 November, highlight improvements to chat performance, image generation, coding agents and training. Long conversations should stream and scroll more smoothly, while code blocks and extended “thinking” panels can scroll independently. The release also adds HTML and PDF chat exports and lets users queue image generations.

For Mac users, the notes say Claude Code and Codex can run with MLX models through unsloth start. They also describe lower VRAM use for FLUX.2 and FLUX.3 image decoding, and say training’s model.generate() is now two to eight times faster by default on T4 GPUs across FastModel models.

Why it matters

These changes touch the daily friction points of local AI work: waiting through long chats, juggling image jobs, moving conversation records elsewhere and getting coding agents running on Mac. The reported training speed-up and lower image-decoding memory use could also matter to people working within tighter hardware limits.

The release notes give the project’s account of its own changes and performance claims. The two-to-eight-times figure is a broad project-reported result, not a guarantee for every workload. Still, this is more than a coat of paint: several improvements have a clear user or operator payoff.

Our read

This release makes Studio look more like a practical workbench for local AI, with smoother everyday use alongside new agent and image features. If you use it, the Mac MLX support, long-chat improvements and image queue are the first changes worth trying. The speed claim is worth testing against your own workload before you plan around it; benchmarks have a habit of being less obliging outside their natural habitat.

What to watch

  • Whether the long-chat improvements hold up in extended real-world sessions.
  • Which Mac MLX models work well with Claude Code and Codex through unsloth start.
  • How the reported training speed-up varies across T4 workloads.

Discussion spark: For local AI tools, which matters more to you: smoother everyday workflows, or measurable gains in speed and memory use?

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
#13403

Update

What changed

Unsloth’s release notes also list Qwen3.8-27B Dynamic v3.1 GGUFs for local use. That gives users another model option alongside the release’s Studio changes, rather than leaving the update as an interface polish exercise.

For image work, Unsloth says FLUX.2 and FLUX.3 decoding uses 36–43% less VRAM at the same speed. The notes also say INT8 and FP8 image models now load on older NVIDIA GPUs, including the T4, and that LTX-2.5 FP8 runs on 24 GB GPUs.

The release further adds an Unsloth MCP preset for coding agents, fixes Claude API models without an explicit token limit stopping at 1,024 tokens, and lets Cosmos3-Nano continue video clips from their first or last frames. A release with this many practical additions is beginning to resemble a workbench refresh, not merely a smoother coat of paint.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.917-beta release includes Qwen3.8-27B Dynamic v3.1 GGUFs, reported lower VRAM use for FLUX decoding, broader image-model compatibility, an MCP preset, and additional API and video features.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#13404

Update

What changed

Unsloth’s release notes add several practical details to the Studio update. Ulysses context parallelism now supports Kimi-Linear, DeepSeek, MiniMax and Gemma 4 text models, while video columns are now supported when training vision models.

The notes also say unsloth start can attach to an already loaded Ollama model. For MLX users, prompts longer than the served context now return a context-length error instead of an empty response, giving users a clearer indication of what went wrong.

Other changes include a fix for OpenAI chats failing when earlier replies were sent in parts, and support for ChatGPT reconnects when refresh tokens are reused or revoked. These additions broaden the release beyond its already reported interface, image and coding-agent changes, with useful fixes for people running local models and training workloads.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s release notes list added Ulysses context-parallelism support for Kimi-Linear, DeepSeek, MiniMax and Gemma 4 text models; video-column support for vision-model training; Ollama attachment support; clearer MLX context errors; and fixes for OpenAI chat replies and ChatGPT reconnection.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#13420

Update

What changed

A new Unsloth commit addresses how Studio displays memory on AMD Strix Halo systems. It fixes a display that counted shared system RAM twice in the VRAM total, which could make the available memory figures misleading for people assessing local models.

The change shows the GPU’s carve-out as VRAM when it is at least as large as the shared portion, and lists separately how much system RAM the GPU can still borrow. It also clarifies the GGUF label for llama cpp Vulkan builds and gives an estimated available-memory figure.

The commit describes this as a display-only correction: model-fit budgets and the backend are unchanged. Its tests used a simulated Strix Halo reading rather than a Strix Halo machine, so the change should not be mistaken for a newly validated performance improvement.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: An Unsloth commit describes a display correction for AMD Strix Halo memory figures: it avoids double-counting host-backed shared RAM in VRAM totals, clarifies the llama cpp Vulkan label, and states that model-fit budgets and the backend are unchanged.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#13424

Update

What changed

Unsloth’s release notes say Cosmos3-Nano can now continue video clips from either their first or last frame, adding a new video-generation capability to the Studio release.

That gives users two possible ends of an existing clip from which to carry on. It is a useful addition for building on footage rather than starting each generation from scratch.

The notes establish the feature’s availability, but do not describe its limits or report a performance evaluation. Those details will matter when people try it on real projects.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#13457

Update

What changed

Unsloth’s release notes add a small but useful chat improvement: read-aloud now begins after the first sentence instead of waiting for the entire response. Listeners can get started sooner, rather than waiting for a long answer to reach its full stop.

Chat preferences, time format and spell check now sync with a user’s account. That should make those settings less of a repeated chore when moving between setups.

The notes also say Unsloth fixed black image outputs with FLUX.3. A practical repair, and a better outcome than asking users to admire a blank canvas.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.917-beta release notes say read-aloud starts after the first sentence, chat preferences, time format and spell check sync with the user’s account, and black image outputs with FLUX.3 have been fixed.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#13474

Update

What changed

Unsloth’s latest Studio change summarises older turns in long cloud chats instead of simply dropping them from the model’s context. The project says this applies when a connection lacks provider-side compaction and a chat passes its context threshold.

Studio asks the same connection and model to summarise earlier turns, keeps recent messages verbatim, and sends both the summary and recent conversation. The full transcript remains in the thread, while conversation recall can supply exact archived passages when a later question needs them.

The summary is stored per thread and reused, rather than regenerated with every message. If summarisation fails, Studio falls back to its previous behaviour. Unsloth says the first compaction adds an extra model call, billed on the same connection, so keeping the chat’s thread intact comes with a small practical cost.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth says Studio now summarises older turns in long cloud chats when a connection does not compact them server-side, preserving recent turns and the full transcript; a failed summary falls back to the previous behaviour, and the first compaction adds a billed model call.

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.