Qwen3.8-Flash-Next + GLM-5.3-Flash

Qwen3.8-Flash-Next and GLM-5.3-Flash can now run locally in Unsloth!

• Run Qwen3.8-Flash on 75GB RAM, GLM-5.3-Flash on 102GB RAM+VRAM
• 5x Faster inference for RAM offloading
• "Infinite" repeated compaction now works
• 100+ chat, reliability and performance improvements

Qwen Guide: https://unsloth.ai/docs/models/qwen3.8-next
Qwen GGUFs: https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF
GLM Guide: https://unsloth.ai/docs/models/glm-5.3-flash
GLM GGUFs: https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF

Highlights

• Qwen3.8-Flash-Next on 75GB RAM
• GLM-5.3-Flash on 102GB total memory
• Smarter GPU + RAM offloading - run larger models with less setup
• Chats recover after disconnects instead of losing the reply
• See what fits before loading with clearer memory estimates

Qwen3.8-Flash-Next

Qwen3.8-Flash-Next is a new 125B multimodal reasoning model and an early preview of Qwen4's architecture.

• The 1-bit Unsloth Dynamic GGUF runs on 75GB RAM or unified memory.
• It's 79% smaller than BF16 while retaining 80% top-1 accuracy.
• Chat with text and images using up to 262K context.
• Switch between None, Low, Medium and Extra High reasoning.
• Preserved Thinking keeps reasoning consistent across longer chats.

GLM-5.3-Flash

GLM-5.3-Flash is Z.ai's new 320B multimodal model, with only 18B parameters active at a time.

• Run the 1-bit model on 102GB of combined RAM + VRAM.
• Chat with text, images and long documents using up to 1M context.
• Switch between Low, High and Max reasoning.
• Stronger coding, agent and vision performance than GLM-5.2.
• Recommended settings are applied automatically in Unsloth.

Chat + tools

• Local chats resume after a disconnect instead of losing the reply.
• Deep Research keeps going when a provider asks it to slow down.
• Vision chats now handle multiple images properly.
• Images returned by MCP tools appear directly in chat.
• Export chats as JSONL for backups or use in other tools.
• Adjust Auto Compaction for longer chats, or turn it off.
• Collapse tool activity by default for cleaner agent chats.

Models + performance

• Large GGUFs automatically split across GPU and system RAM.
• See estimated memory usage before loading a model.
• View VRAM usage directly from your downloaded models.
• Model settings stay saved when switching chats.
• Search and download embedding models directly from Hugging Face.
• Text-to-speech models only load when you actually use them.

Desktop + reliability

• Linux voice recording fixed.
• NVIDIA + Wayland interface freezes fixed.
• AMD model loading crashes fixed.
• llama.cpp models now load from Windows profiles with non-English characters.
• Non-English web links now work properly as chat sources.
• Desktop download links always point to the latest stable release.

What's Changed

• Bump install.sh / install.ps1 pin to unsloth>=2026.8.21 by @danielhanchen in #9699
• fix(studio): resolve PowerShell by absolute path in the update gate by @yzxcj797 in #9452
• Prevent shared preview loads from evicting the active Studio model by @NilayYadav in #7104
• studio: harden lockfile audit followups for #5604 by @danielhanchen in #5695
• Fix Linux voice recording by capturing raw PCM where WebKitGTK's MediaRecorder produces no audio by @Fizza-Mukhtar in #9564
• Studio: take remend 1.3.1, which stops repairing markdown that is already complete by @danielhanchen in #9667
• CLI: follow model reasoning default in unsloth start by @shimmyshimmer in #9733
• Name the encoding when the poll probe writes its artifact by @danielhanchen in #9691
• studio: add a freeze report script for Linux desktop interface hangs by @danielhanchen in #9668
• Fail the mac tab-capability smoke on a backend that never comes back, not on a stall by @danielhanchen in #9711
• Warm the app before timing /api/health, instead of widening the ceiling by @danielhanchen in #9724
• docs: remove duplicate Search bullet and fix error messages by @lxcxjxhx in #9566
• Follow up #9435: restore the rest of the moved names, and read quant labels through the variant lister by @danielhanchen in #9701
• Studio: use --load-mode none when the model fits without paging by @danielhanchen in #9565
• Studio: spill weights with -ot instead of layers, so the KV cache stays on the GPU by @danielhanchen in #9675
• Reject a negative chunking stride instead of silently dropping text by @vineethsaivs in #9249
• Studio: fix web pages with non-English URLs failing to load by @NilayYadav in #9721
• Studio: align composer and research activity icons by @shimmyshimmer in #9755
• desktop: turn webkit compositing off on the one wayland nvidia setup that freezes by @danielhanchen in #9734
• studio: keep a resumed turn's provider metadata through the continuation merge by @danielhanchen in #9684
• Studio: render MCP images returned as embedded resources by @NilayYadav in #9636
• Studio: send the most recent chat image to vision models by @NilayYadav in #9638
• Studio: align sidebar list header actions, and mark downloaded models with a dot by @shimmyshimmer in #9740
• Prevent stale hydration from overwriting remembered model settings by @alkinun in #9767
• Handle invalid Krea2 model indexes clearly by @alkinun in #9760
• Studio: track API token usage in Profile by @Etherll in #9320
• Studio: tell the model which documents are attached by @NilayYadav in #9718
• Studio: use live trending GGUFs in agent setup by @shimmyshimmer in #9752
• studio: preserve gguf settings across chat switches by @mahiatlinux in #9562
• studio: add message jsonl export by @mahiatlinux in #9614
• Validate studio_install_id before baking it into the launcher by @danielhanchen in #9741
• studio: show lm studio model names in chat by @mahiatlinux in #9556
• Studio: fix per-chat settings and fork counts for app-created chats by @NilayYadav in #9639
• studio: show the train run preview as a column on a laptop window by @mahiatlinux in #9762
• Use Unsloth for bare Studio product references by @shimmyshimmer in #9732
• Studio: stop the temporary toggle tagging a saved app-created chat by @danielhanchen in #9771
• Studio: honour forced tool_choice on local GGUF tool loops by @Souravrajvi0 in #9773
• Desktop: publish stable latest downloads by @oobabooga in #9652
• Give every installing job its own pip cache key by @danielhanchen in #9781
• Studio: report confirmed explicit GGUF context by @Doud-FR in #9713
• Studio: keep Deep Research alive when the provider asks it to slow down by @NilayYadav in #9720
• Studio: don't silently drop extra images from a message by @NilayYadav in #9719
• Studio: hide model disclaimer by default by @shimmyshimmer in #9789
• README: drop the Linux (Arm64) download row, which serves a macOS file by @danielhanchen in #9783
• studio: newline-terminate exported jsonl records by @mahiatlinux in #8759
• Take pytest tests/ from 16m22s to 4m17s at -n 16 by @danielhanchen in #9794
• Derive the fp8 fbgemm test tolerance from the bf16 ULP by @danielhanchen in #9795
• Make importing unsloth.models._utils twice in a process survive by @danielhanchen in #9793
• Studio: let the model decide when a message needs Deep Research by @NilayYadav in #9726
• Fix GGUF mmproj pairing for derivative repos by @Imagineer99 in #6574
• Remove duplicate TRL vLLM generation patch by @alkinun in #9764
• Studio: widen what the offload planner can place instead of declining by @danielhanchen in #9778
• Studio: turn the offload planner on by default by @danielhanchen in #9779
• studio: add a sampling seed to the chat run settings by @mahiatlinux in #9144
• Return no base model, not the bare unsloth/ org, for a timestamp-less run directory by @rajarshidattapy in #9753
• Fix RAG fallback for oversized chat attachments by @Supersynergy in #9630
• Add Estimated Memory Usage to the Load Model panel by @shimmyshimmer in #9525
• Build sd.cpp test fixtures from _binary_names() instead of the POSIX literal by @derrickrajkumar10 in #9707
• Fix Sentence-Transformers 6 model_forward_params None handling by @kozistr in #9786
• Skip an Ollama manifest whose JSON is not an object, instead of 500ing the model picker by @rajarshidattapy in #9754
• Studio: auto-scroll the startup and update logs by @Lwrless in #9806
• Collect the pip-cache naming guard in the one job a workflow-only PR starts by @danielhanchen in #9802
• Fix main: assert the tool loop controller catalog structurally, not by source proximity by @danielhanchen in #9809
• Studio: add the four settings.agents keys to every locale overlay by @danielhanchen in #9808
• Studio: bound fence tokenization by line length, not by the wall clock by @danielhanchen in #9787
• studiobench: a null-control refusal that read the settled thread is an observation, not a blank by @danielhanchen in #9800
• Studio: warm fence grammars on real text, not an empty string by @danielhanchen in #9799
• Studio: let maths-bearing blocks take containment, behind a flag by @danielhanchen in #9731
• fix(studio): prevent OXC validator subprocess hangs by @Biotrioo in #9759
• Studio: stop chat from auto-loading a TTS model by @NilayYadav in #9813
• Fix llama.cpp loading for Unicode Windows profiles by @Imagineer99 in #8852
• fix(studio): distinguish dedicated vram from shared gpu memory (#9242) by @yzxcj797 in #9247
• Studio: clear the old transcript when a transcription fails by @NilayYadav in #9816
• Studio: embedding model picker searches the Hub and downloads what you pick by @shimmyshimmer in #9739
• Define "where local models live" once, so a BOM cannot split the answer in two by @Biotrioo in #9757
• Studio: refuse keyless access to a cross-site browser request by @danielhanchen in #9683
• Studio: resume local chat generation after disconnect by @Etherll in #9187
• Studio: Show the model's thinking on the preview page by @NilayYadav in #9815
• Studio: restore menu focus after outside dismissal by @alkinun in #9772
• Studio: show a VRAM memory bar on downloaded models by @IMJONEZZ in #7880
• Studio: fix the AttributeError that 500s every planned GGUF load by @danielhanchen in #9821
• Studio: tighten the comments added by the memory estimate PR by @danielhanchen in #9825
• Fix the studio PEP 604 union ratchet and name what breaches it by @danielhanchen in #9826
• Studio: keep the memory row readable on a narrow panel by @danielhanchen in #9824
• route thinking controls to ollama by @Souravrajvi0 in #9770
• Studio: read the memory row's captions past their line-breaking glue by @danielhanchen in #9831
• Studio: persist model disclaimer across instances by @shimmyshimmer in #9835
• Studio: consolidate the two memory-estimate stacks by @danielhanchen in #9830
• fix(prompt storage): rework prompt storage into a master-detail layout by @LeoBorcherding in #8692
• Studio: add Chat settings to disable or ease GGUF auto-compaction by @Souravrajvi0 in #9774
• Stop the slot re-fit tests encoding the fit floor as a literal by @danielhanchen in #9845
• Stop a left-padded row that attends to nothing returning NaN (#9708) by @danielhanchen in #9848
• Reject a macOS llama.cpp prebuilt that dyld will not load by @danielhanchen in #9843
• Fix the dev-server white screen from a two-barrel import cycle by @danielhanchen in #9852
• Stop a resumed turn retracting text it already streamed by @danielhanchen in #9851
• Studio: stop the dense torchao quant probe from crashing the backend on AMD ROCm by @oobabooga in #9828
• Stop a generation recovery rewinding the reply it is following by @danielhanchen in #9853
• Stop a recovery following a run this tab is already streaming by @danielhanchen in #9854
• Queue a parked send when a run started while it waited by @danielhanchen in #9810
• Add setting to collapse tool activity by default by @Imagineer99 in #9803
• fix(studio): restore parallel vision chat slots by @Biotrioo in #9842
• fix keyless access before desktop password setup by @mahiatlinux in #9855
• Fix dill pickling whole modules by value on an off-prefix install by @danielhanchen in #9791
• Studio: stop a small context window ending a tool-using turn with nothing by @danielhanchen in #9768
• Studio: warn instead of refusing to load a model bigger than VRAM plus RAM by @danielhanchen in #9822
• Apply the kwarg-spacing hook to the no-progress tool results test by @danielhanchen in #9860
• Keep the context length a load was given instead of resetting it to Auto by @danielhanchen in #9841

New Contributors

• @Doud-FR made their first contribution in #9713
• @rajarshidattapy made their first contribution in #9753
• @Supersynergy made their first contribution in #9630
• @derrickrajkumar10 made their first contribution in #9707
• @kozistr made their first contribution in #9786
• @Lwrless made their first contribution in #9806
• @IMJONEZZ made their first contribution in #7880

Full Changelog: v0.1.803-beta...v0.1.804-beta