Meta Muse Glimmer

Meta has released Muse Glimmer, the first open model from Meta Superintelligence Labs. Muse Glimmer is a dense, 30B parameter model by Meta, built for local agentic and coding workflows. Released under the Apache 2.0 license you can run Muse Glimmer 30B via Unsloth and Unsloth Dynamic quants for the best performance.

Run Muse Glimmer
Fine-tune Muse Glimmer

Muse Glimmer 30B can run locally on 20GB RAM/VRAM setups, including Mac and GPUs. You can also train and fine-tune Muse Glimmer 30B using Unsloth on 20GB VRAM.

Updating / installing Unsloth

macOS, Linux, WSL:

curl -fsSL https://unsloth.ai/install.sh | sh

Windows:

irm https://unsloth.ai/install.ps1 | iex

What's Changed

• Studio: judge the cached-pipeline signals on the snapshot the row actually loads by @danielhanchen in #7851
• Bump install.sh / install.ps1 pin to unsloth>=2026.8.3 by @danielhanchen in #7860
• Studio: make the sidebar menu separators visible in dark mode by @shimmyshimmer in #7865
• Studio: run sandbox matplotlib headless by @NilayYadav in #7789
• Don't send desktop startup requests to an unvalidated port 8888 by @NilayYadav in #7852
• Support independent prompt queues across parallel chats by @Imagineer99 in #7629
• add a background image to the macos dmg install window by @mahiatlinux in #7827
• Studio: keep Video under More until a user pins it by @shimmyshimmer in #7863
• Studio: show generated media unrounded, and fix the selected thumbnail by @shimmyshimmer in #7883
• Studio: drop the New badge from the Video nav row by @shimmyshimmer in #7890
• Studio: stop a clicked tooltip from hanging over dialogs by @shimmyshimmer in #7836
• Desktop: move update controls to Settings > General by @oobabooga in #7807
• Fix the stale response-model label assertion that has main red by @danielhanchen in #7874
• Studio: detect stalled Deep Research output streams by @alkinun in #7839
• Guard the release-desktop version input and document the Cargo version by @danielhanchen in #7889
• Studio: add Qwen3-ASR dictation, cancellable downloads, and a download prompt by @shimmyshimmer in #7835
• Studio: even out the gap around the More flyout's rule by @shimmyshimmer in #7899
• Studio: stop the New chat row looking hovered on a blank new chat by @shimmyshimmer in #7898
• Studio: fix remaining translation defects after #7797 by @danielhanchen in #7850
• Studio: fix download card alignment and surface the stop button by @shimmyshimmer in #7848
• Keep the file-edit turn 2 prompt to one instruction, and stop waiving its cap by @danielhanchen in #7846
• Studio: align Tauri sidebar spacing by @wasimysaid in #7901
• Fix sidebar bottom fade cutting off the last row mid-glyph by @shimmyshimmer in #7902
• Bump install.sh / install.ps1 pin to unsloth>=2026.8.4 by @danielhanchen in #7906
• Answer GGUF variants from the cache the request names, and skip an unreadable one by @danielhanchen in #7857
• Desktop: warn before quitting during an update and skip installer fallback on cancel by @NilayYadav in #7793
• Desktop: publish drafts without rebuilding by @wasimysaid in #7907
• Studio: restore locale parity and gate it in CI by @danielhanchen in #7904
• Studio: make training and model-picker copy name UI that exists by @danielhanchen in #7905
• fix(studio): run bash and name the real platform for the terminal tool on Windows by @LeoBorcherding in #7885
• Desktop: fix Linux release asset names by @wasimysaid in #7911
• Unsloth Studio: keep the Images and Video headers clear of the custom titlebar by @LeoBorcherding in #7918
• fix(studio): keep child spawns alive under a pid-1 parent by @mahiatlinux in #7894
• Ask uv for a Python that can import torch, and skip the one that cannot by @danielhanchen in #7830
• Improve desktop startup status by @wasimysaid in #7924
• Tauri smoke: stop uploading the whole target/debug tree (350GB of artifacts) by @danielhanchen in #7930
• Restore warnings_issued for trl trainers on transformers 5.1+ by @danielhanchen in #7854
• Fetch a sentence-transformers model's subfolder weights instead of pruning them by @danielhanchen in #7855
• Do not enter CUDA autocast for a float32 model during generate by @danielhanchen in #7900
• Locate the dictation model search box by test id, not by placeholder copy by @danielhanchen in #7909
• Record a stopped repair update as canceled rather than failed by @danielhanchen in #7910
• Bound dataset_num_proc by memory, and stop treating 1 as "no multiprocessing" by @danielhanchen in #7831
• Studio: quiet the image and video generation progress polls by @danielhanchen in #7869
• Studio: stop the Images Train settings columns overlapping by @danielhanchen in #7895
• Fake the Windows blocklist in the mangled-switch test so it runs off Windows by @danielhanchen in #7934
• Refine desktop media header alignment by @wasimysaid in #7859
• Read the GGUF context length from the snapshot that answered, not the repo cache dir by @danielhanchen in #7920
• Tighten remaining artifact uploads (follow-up to #7930) by @danielhanchen in #7939
• Studio: prevent long streaming Markdown stalls by @oobabooga in #7892
• Fix studio server crash when launched without a console (Windows) by @legobele in #7932
• Fix GGUF export writing outside the chosen save directory on Windows (#7897) by @danielhanchen in #7937
• Fix Windows PowerShell resource path authorization by @Imagineer99 in #7943
• Desktop: fix the mobile-layout fallout of Windows text scaling by @danielhanchen in #7927
• Desktop: make the close button quit outside macOS by @danielhanchen in #7929
• Desktop: pin the shipped backend package version so a shared venv gets upgraded by @danielhanchen in #7926
• Studio: screen the detached-window test against the Windows blocklist too by @danielhanchen in #7936
• Require every declared denoiser, and every shard of it, before a cached pipeline reads as complete by @danielhanchen in #7871
• Fix the Windows-only snapshot path assertion in the diffusion prefetch test by @danielhanchen in #7933
• Studio: paint the image and video pickers from the curated catalog by @danielhanchen in #7896
• Bump install.sh / install.ps1 pins to unsloth>=2026.8.5, unsloth-zoo>=2026.8.4 by @danielhanchen in #7967
• Studio: consistent rounded hover boxes for nav icon buttons by @shimmyshimmer in #7970
• Studio: lead the Context Length note with the automatic fit by @shimmyshimmer in #7988
• Studio: test the Windows launcher shape, not just the path normalizer by @danielhanchen in #7969
• Studio: stop model rows staying lit after the dots menu closes by @shimmyshimmer in #7987
• Update the nav icon button size contract to the 30px box by @danielhanchen in #7979
• Studio: fetch the gated diffusion bases from ungated unsloth mirrors by @danielhanchen in #7952
• Studio: fix training stuck at 100% when the worker will not exit by @danielhanchen in #7938
• Report the real cause when a save or GGUF export fails by @danielhanchen in #7861
• Locate the dictation model search box by test id, not by its copy by @danielhanchen in #7914
• CI: drop three macOS legs that assert nothing a Linux runner does not by @danielhanchen in #7974
• CI: fold the mac API smoke into the UI smoke, drop the subsumed install leg by @danielhanchen in #7975
• Installer: preserve rollback state and merge on partial venv move (#7810) by @namdamdoi68-oss in #7891
• CI: fold the mac update smoke into the same job, running last as teardown by @danielhanchen in #7980
• CI: save caches only on main, so PR runs stop evicting the caches they need by @danielhanchen in #7990
• Studio: drop the streaming animation wrappers once a message completes by @danielhanchen in #7949
• Skip a test module with a missing optional dep instead of exiting pytest by @danielhanchen in #7971
• Studio: restore the MLX gradient-norm chart by matching CUDA's clipping default by @Lyxot in #7917
• Studio: stop a GGUF pick from pulling the wrong companion bytes by @danielhanchen in #7870
• Studio: rework the media pane headings and tidy the Create controls by @shimmyshimmer in #7986
• Studio: keep the sidebar fade in sync on resize and centre the profile row by @shimmyshimmer in #7977
• CI: run the three mac inference smokes as phases of one job by @danielhanchen in #7978
• Studio: link the Train page across to image training by @shimmyshimmer in #7996
• fix(studio/install): keep the rollback reference when a venv move stops partway (#7810) by @Anai-Guo in #7823
• Studio: format profile stat numbers with the chosen locale by @danielhanchen in #7916
• Settle deferred compile-mode switches between training steps by @danielhanchen in #7955
• Studio: simplify desktop setup progress UI by @wasimysaid in #8011
• Tighten the gated-base preflight comments by @danielhanchen in #8003
• CI: point rust-cache at the real v2.9.1 tag and label the unlabelled pins by @danielhanchen in #8005
• CI: move macOS runners off the retiring macos-14 image by @danielhanchen in #8004
• Studio: fetch the sd.cpp companion assets from unsloth mirrors, not community repacks by @danielhanchen in #7995
• Studio: make the fp8 smoke probe catch a missing activation scale floor by @danielhanchen in #8008
• Diffusion: four fixes salvaged from the stale #7021 by @danielhanchen in #8006
• studio: persist the settings dialog picks that reset on every tab switch by @mahiatlinux in #7991
• Studio: bound every teardown await so a swallowed cancel cannot wedge the API by @mahiatlinux in #7881
• feat(studio): rework train page setup flow by @Sneakr in #7633
• Give the Windows sandbox bash a userland, and honour TEMP/TMP by @danielhanchen in #8009
• Screen the cmd spellings the blocklist lexer quotes by @danielhanchen in #8010
• CI: extract the POSIX local install into a composite action by @danielhanchen in #8015
• Add option to launch Studio without opening the default browser by @shimmyshimmer in #7016
• Stop the rapid-submit chat UI check failing on a fast runner by @danielhanchen in #7972
• CI: run the three Windows inference smokes as phases of one job by @danielhanchen in #8013
• fix(studio/rocm): don't install for a shadowing iGPU on mixed AMD hosts (#7776) by @Anai-Guo in #7778
• Diffusion: replace the attention perf claims with re-measured numbers by @danielhanchen in #8021
• Desktop: fix Linux AppImage compatibility on newer distributions by @oobabooga in #7953
• Model hub: report why Hugging Face is unreachable instead of claiming you are offline by @danielhanchen in #7829
• Studio: stop two backend tests from asking the host what the answer is by @danielhanchen in #8022
• Studio: do not let the default tool policy reject every plain Anthropic message by @danielhanchen in #8023
• Fix desktop window chrome by @wasimysaid in #8025
• Order the Mac inference phases so the gates stop fighting each other by @danielhanchen in #8024
• Extract the API-only Studio boot into one script by @danielhanchen in #8029
• Fix the desktop clean-machine CI: asset patterns and the retired #7547 pins by @danielhanchen in #8026
• Studio: center sheet header close buttons by @oobabooga in #7957
• feat(studio): add settings-managed remote access by @Lyxot in #7875
• Studio: add opt-in DSpark speculative decoding by @oobabooga in #7968
• Bump install.sh / install.ps1 pin to unsloth>=2026.8.6, unsloth-zoo>=2026.8.5 by @danielhanchen in #8038
• fix studio responsiveness across desktop and mobile web by @mahiatlinux in #8016
• Revert "Add option to launch Studio without opening the default browser" by @danielhanchen in #8040
• Bump install.sh / install.ps1 pin to unsloth>=2026.8.7 by @danielhanchen in #8041
• Studio: send the GGUF filename when a collapsed On Device row is picked by @oobabooga in #8053
• Validate an explicit GGUF variant before unloading the resident model by @NilayYadav in #7862
• fix(studio): stop showing manifest-less cache entries as downloading (#7858) by @Anai-Guo in #7864
• release-desktop: skip the Defender scan when Defender is unavailable by @danielhanchen in #8059
• Report min_length as 0 when no sample has content by @vineethsaivs in #7872
• Stop the health watchdog bypassing the backend startup grace period by @NilayYadav in #7958
• Studio: honor MLX KV cache quantization and chat template overrides by @Lyxot in #8007
• test: isolate the no-version gfx-override test from a real /opt/rocm by @danielhanchen in #7397
• Studio: stop /api/inference/status raising on every loaded GGUF by @danielhanchen in #8074
• CI: share the gemma-3-270m-it model cache across Linux, macOS and Windows by @danielhanchen in #8068
• CI: share the Qwen3.5-2B tool-calling GGUF cache across Linux, macOS and Windows by @danielhanchen in #8069
• Fix the four Studio reds on main by @danielhanchen in #8072
• Studio: harden DSpark/MTP drafter detection and companion reclaim by @danielhanchen in #8037
• Studio: keep the stdio MCP session alive when a tool call is cancelled by @NilayYadav in #7625
• Reuse the torch2.10 prebuilt accelerator wheels on torch 2.12 by @danielhanchen in #7495
• Studio: line up every sidebar row pill with New Chat by @shimmyshimmer in #8079
• Normalize the /api/health wait and extract it into one script by @danielhanchen in #8060
• Make a fatal X11 error visible in the Linux desktop legs by @danielhanchen in #8063
• Studio: call XInitThreads so the Linux desktop app stops exiting on startup by @danielhanchen in #8075
• Studio: bound Deep Research first output waits by @alkinun in #7998
• Allow current MLX dependencies with valid Studio sidecars by @Lyxot in #7061
• Desktop: keep slide-in panels below the custom titlebar by @oobabooga in #7954
• CI: stop running workflows on the pip branch, and mirror push path filters for the non-Studio workflows by @danielhanchen in #8085
• CI: always run Studio export/load-orchestrator and the security audit by @danielhanchen in #8086
• release-desktop: cache trusted-signing-cli and add job timeouts by @danielhanchen in #8087
• Show the public model id in the selector when the loaded model isn't in the picker list by @LeoBorcherding in #7966
• Explain why the gated-deltanet fast kernels were disabled by @danielhanchen in #8073
• Studio: fix the Windows desktop setup log mojibake and double-printed steps by @danielhanchen in #8083
• Installer: keep status messages off the progress bar line by @oobabooga in #8052
• Fix the status runtime-fields contract test left stale by #8074 by @shimmyshimmer in #8088
• Studio: pin the /status chat template echo behaviour by @shimmyshimmer in #8080
• Desktop: Fix unresponsive model selector on Images/Video pages on macOS by @NilayYadav in #8048
• Fix Swagger UI and ReDoc being blanked by the studio CSP by @NilayYadav in #8042
• Stop referencing a TAURI signing password secret that does not exist by @danielhanchen in #8095
• CI: add an inert cross-repo capacity sweeper (do not merge yet) by @danielhanchen in #8071
• tests: extend the version-compat matrices to the current PyPI releases by @danielhanchen in #8020
• Scope the security suite's offline guard to the tests that want it by @danielhanchen in #8054
• Compare the drafter path with forward slashes on Windows too by @danielhanchen in #8056
• Desktop: start at login, hidden in the menu bar by @NilayYadav in #8033
• CI: stop main and nightly runs cancelling each other by @danielhanchen in #8070
• Studio: stop the macOS launcher killing a warming backend, and the Mac tab blackout by @danielhanchen in #8076
• Security: baseline seven reviewed scan_packages findings by @danielhanchen in #8096
• Studio: label the chat model bar by repo, not the HF cache path by @danielhanchen in #8090
• Security: pin the credential-send allowlist entry to the reviewed file by @danielhanchen in #8104
• CI: move the capacity sweeper off the retiring ubuntu-22.04 image by @danielhanchen in #8105
• release-desktop: add a VirusTotal pre-flight scan of the release bundles by @danielhanchen in #8089
• release-desktop: fail closed when trusted-signing-cli is missing or broken by @danielhanchen in #8106
• Desktop: avoid WebKitGTK dmabuf crashes on Wayland by @wasimysaid in #8028
• Studio: cap llama.cpp source build parallelism by RAM by @shimmyshimmer in #8061
• Match the LM Studio GGUF suffix case insensitively by @danielhanchen in #8100
• Bump the actions group across 1 directory with 11 updates by @dependabot[bot] in #8101
• Guard Windows Studio installs against active runtimes by @Etherll in #7764
• Desktop: prevent thin AppImage environment failures by @oobabooga in #8055
• Studio: preserve the Windows launcher during updates by @Etherll in #8092
• Desktop: fit windows within the monitor work area by @oobabooga in #8047
• Studio: keep dictation status reads off the sidecar lock by @shimmyshimmer in #8017
• Desktop: Fix "Copy path" failing in the desktop app by @NilayYadav in #7959
• Uninstall: remove WebView runtime data so reinstalls don't serve a stale frontend by @shimmyshimmer in #7360
• Fail CI on new yes/no prompts in the installers by @danielhanchen in #8094
• Studio: continue a response that stopped early by @shimmyshimmer in #8064
• Studio: say what a failed launcher move-aside costs by @danielhanchen in #8109
• Keep up with transformers 5.x, TRL 0.22.x and peft by @danielhanchen in #7866
• Setup: clear stale WebView caches on install and update, keep user settings and data by @shimmyshimmer in #7361
• Studio: find models already on the device before downloading, and make the first download cancellable by @shimmyshimmer in #7379
• [Bug] Unsloth Studio: let fenced HTML canvases honor the canvas network-access setting by @LeoBorcherding in #8098
• Studio: keep chat titles whole so a wider sidebar shows more of them by @shimmyshimmer in #8078
• Studio: keep the composer tool pills on one line by @shimmyshimmer in #8111
• Preflight the Codex GGUF check before loading the model by @NilayYadav in #7873
• Studio: keep the upgrade intent when the pip fallback runs by @danielhanchen in #8112
• Improve desktop installer progress UX by @wasimysaid in #8102
• studio: expose llama-server batch and micro-batch sizes as first-class load settings by @mahiatlinux in #7973
• Bump install.sh / install.ps1 pin to unsloth>=2026.8.8 by @danielhanchen in #8114
• Clamp Deep Research max_tokens to loaded context window by @Souravrajvi0 in #7985
• Desktop: block HTTP loopback requests from untrusted images by @oobabooga in #8046
• Studio: keep inference status probes off the event loop by @oobabooga in #7944
• Add Model memory settings to keep a loaded model in VRAM by @shimmyshimmer in #8002
• perf(studio): make local model inventory scans linear by @Lyxot in #8043
• Repair seven tests left behind by their own subjects by @danielhanchen in #8132
• fix(security): re-baseline the huggingface-hub HTTP backoff loop for 1.x by @danielhanchen in #8135
• Studio: check llama.cpp cache access before setup by @oobabooga in #8032
• studio: fix deep research progress reporting, loopback auth and model retries by @mahiatlinux in #8129
• Reserve host RAM for AMD unified-memory APUs on the ROCm probe by @danielhanchen in #8131
• fix(AMD): let diffusion load on Windows ROCm, and say what a gated repo needs by @LeoBorcherding in #7981
• Desktop: ensure backend ownership for Tauri installs by @oobabooga in #8034
• Studio: keep row actions clear of overlay scrollbars by @oobabooga in #8031
• Report a denied UNSLOTH_LOCAL_LLAMA_CPP_DIR instead of aborting on errexit by @danielhanchen in #8143
• Studio: per-request perf stats and slot usage in the API monitor by @NilayYadav in #8045
• fix(studio): harden training setup, lifecycle, and audio loading by @Sneakr in #8103
• Pin the accelerator gates in the DiT precision preflight test by @danielhanchen in #8155
• Studio: fix STT download progress and cache safety by @oobabooga in #8035
• Fix desktop image drops for chat attachments by @Souravrajvi0 in #7982
• Studio: fall back to CPU when auto-selected Vulkan crashes by @oobabooga in #8050
• Studio: stop research threads 409-ing on autosave, keep server-managed messages intact by @NilayYadav in #7956
• Studio: retry Xet once on a stall before falling back to HTTP by @danielhanchen in #8166
• Deduplicate the scan-packages baseline, and gate it with tests by @danielhanchen in #8110
• Add Markdown conversation export by @ousamabenyounes in #7664
• Bump the cargo-tauri group across 1 directory with 18 updates by @dependabot[bot] in #8107
• CI: prune superseded Actions caches by @danielhanchen in #8167
• Publish studio_install_id without clobbering a concurrent writer by @danielhanchen in #8145
• Desktop: refuse to republish an existing release version by @oobabooga in #7941
• Studio: surface API errors and align audio limits by @oobabooga in #7946
• Stop the CUDA test spoofs reporting an exhausted card by @danielhanchen in #8123
• Keep download scopes out of the GGUF quantization list by @shimmyshimmer in #8153
• Do not choose flex attention on a card that cannot run its kernel by @danielhanchen in #8137
• Render a static prefix when default_system_message is None by @vineethsaivs in #8117
• Studio: rate the API monitor Throughput tile on the decode window by @shimmyshimmer in #8139
• [Fix] Unsloth Studio: Strip ANSI escape codes from Studio tool output panes by @Souravrajvi0 in #7976
• Attach save_lora without a vLLM engine by @danielhanchen in #8127
• Studio: bring back the sidebar scrollbar without losing row width by @shimmyshimmer in #8147
• Studio: exempt shared top-level trees from the integrity check for our own wheels too by @danielhanchen in #8179
• Do not stage diffusion files that are already in the cache by @shimmyshimmer in #8154
• Catch the llama-prebuilt validator stubs up with rocm_gfx by @danielhanchen in #8182
• Pin the auto-load success toast by its branch, not by one spelling by @danielhanchen in #8183
• Studio: route every Windows installer line through the UTF-8 stdout sink by @danielhanchen in #8148
• Studio: hold the Mac capability verdict while an MLX repair can still overturn it by @danielhanchen in #8152
• Count a RECORD path claim before the row is filtered out by @Imagineer99 in #8181
• Fix on-device GGUF image/video picks asking for a quantization by @shimmyshimmer in #8149
• Enforce the interface's video resolutions in the API too by @danielhanchen in #8159
• Studio: add a loaded models indicator with per-model eject by @shimmyshimmer in #8082
• Desktop: show a closing overlay while the backend shuts down by @danielhanchen in #8151
• Studio: make the gallery video link absolute so the desktop app can play and save it by @danielhanchen in #8173
• Bump install.sh / install.ps1 pin to unsloth>=2026.8.9 by @danielhanchen in #8189
• Studio: eliminate Cargo and Clippy warnings by @wasimysaid in #8192
• studio: serve the deep research event stream over post so proxies stop buffering it by @mahiatlinux in #8205
• Desktop release: build the three platform legs in parallel by @danielhanchen in #8191
• Desktop release: start the publish job alongside the build matrix by @danielhanchen in #8193
• tests: make pytest tests collect cleanly and drop the deleted QVQ registry mirror by @danielhanchen in #8206
• Desktop release: run the VirusTotal scan after publishing, in its own job by @danielhanchen in #8194
• tests: build the hub-error fixtures without calling init (huggingface_hub 1.x) by @danielhanchen in #8207
• Studio: stop MTP forcing llama-server to a single parallel slot by @danielhanchen in #8172
• test(studio): stop the backend suite depending on a live Hub by @danielhanchen in #8174
• Repin six UI contracts on what they guard, not how it was spelled by @danielhanchen in #8197
• Studio: only dodge an overlay box that reaches the corner stack by @shimmyshimmer in #8210
• Studio: surface an MLX completion-masking miss as a warning, not a status line by @BardiaKoopah in #8133
• Desktop: block accidental reloads in the Windows WebView by @oobabooga in #7945
• Studio: preserve training config values across YAML load and save by @oobabooga in #7942
• Stop a finished download reporting 0 B and offering Retry by @danielhanchen in #8160
• Install a CUDA-matched xFormers on Windows instead of whatever PyPI serves by @danielhanchen in #8156
• fix(studio): release the inference subprocess's VRAM when its last model unloads by @LeoBorcherding in #8221
• fix(studio): reserve a bounded OS headroom on unified ROCm APUs, not a flat 20% by @LeoBorcherding in #7960
• Studio: fix Windows desktop updates failing integrity checks by @Etherll in #8185
• Report the precision actually in use, and refuse an explicit one we cannot honor by @danielhanchen in #8165
• Pin the training precision, video preset and image metadata contracts by @danielhanchen in #8162
• Say when xFormers is installed but its kernels cannot load by @danielhanchen in #8157
• Revert "Say when xFormers is installed but its kernels cannot load" by @danielhanchen in #8237
• Studio: stream oversized diffusion components on low-VRAM GPUs by @oobabooga in #8141
• fix(studio): wire llama's OpenMP runtime into slim whisper installs by @danielhanchen in #8230
• fix(studio): verify TLS against the OS trust store at runtime (corporate TLS-inspection proxies) by @Cesarsk in #8108
• Tell users to upgrade the dependency, not transformers, when a --no-deps install leaves a floor unmet by @danielhanchen in #8212
• Repo tests: pin the post-publish VirusTotal contract instead of the old pre-flight one by @danielhanchen in #8240
• Let the pipeline gate survive a diffusers whose lazy submodule cannot import by @danielhanchen in #8227
• Stop the idle-unload TTL test racing a loaded runner by @danielhanchen in #8175
• Let the floating monitor paint over the notification stack by @danielhanchen in #8199
• Format the release permissions test to the kwarg-spacing hook by @danielhanchen in #8245
• Format the release-permissions test so pre-commit.ci stops failing on main by @danielhanchen in #8247
• add folder picking and drag-drop for paired image and caption datasets by @mahiatlinux in #8216
• Backfill the conversion symbols peft imports and transformers 5 dropped by @danielhanchen in #8051
• Studio: only clean up GGUF artifacts the export owns by @oobabooga in #7940
• Resume diffusion training from a stop-and-save checkpoint by @danielhanchen in #8163
• Stop padding-free SFT from tripping the TRL >= 1.0.0 max_length guard by @danielhanchen in #7951
• Fix Studio loaded-model GGUF quant label by @Imagineer99 in #8238
• Do not scan vendored packages that refuse to import on our floor by @danielhanchen in #8252
• Desktop: confirm quit during training by @NilayYadav in #7790
• Fetch the llama.cpp validation model only when the smoke test reads it by @danielhanchen in #8248
• Make install.ps1 work with the user's PowerShell profile loaded by @danielhanchen in #8161
• Name a torchvision whose compiled ops do not match torch by @danielhanchen in #8128
• Diffusion: source every image family's companions from unsloth mirrors by @danielhanchen in #8030
• studio: infer splits for metadata-free cached datasets by @Ricardo-M-L in #8144
• Reject an incompatible GGUF and base pair before the download, not after by @danielhanchen in #8158
• Make the desktop release contract tests fail when the contract breaks by @danielhanchen in #8228
• GRPO: wrap the model that owns the head when falling back to hidden states by @danielhanchen in #8231
• Diffusion: allow fp8 on Qwen-Image again, and split the training deny by @danielhanchen in #8018
• Report the network error when the offline retry finds nothing cached by @danielhanchen in #8121
• Keep an explicitly requested float32 model in float32 without bfloat16 by @danielhanchen in #7867
• Do not let a broken TensorFlow install break the Unsloth import by @danielhanchen in #8124
• Carry the sliding window into SDPA, and treat a zero window as no window by @danielhanchen in #8253
• Resolve latest mlx-vlm and route its new processors by @Lyxot in #8251
• Studio: align Images and settings layouts by @wasimysaid in #8164
• Studio: realign the image precision contracts with the fail-closed refusal by @danielhanchen in #8257
• Studio: name the quant a diffusion GGUF pick actually loaded, instead of the compute dtype by @danielhanchen in #8241
• Studio: budget a pre-cast fp8 text encoder at the size it actually loads by @danielhanchen in #8218
• Desktop: keep probing the announced backend port instead of discarding it by @danielhanchen in #8211
• Stop the CPU fake-train tests from spoofing torch for the whole session by @danielhanchen in #8259
• Studio: grey the sidebar dot a finished chat leaves by @shimmyshimmer in #8258
• Studio: stop leaking child processes on an abnormal exit by @danielhanchen in #8170
• Restore SFT loss_type='nll' default over TRL 1.7.0 chunked_nll by @danielhanchen in #8261
• Assert the image family's pipeline class in the training preflight, before the GPU teardown by @danielhanchen in #8236
• Studio: let image generation be stopped by @danielhanchen in #8219
• Studio: make the image memory plan resolution-aware and stop reserving the text encoder during sampling by @danielhanchen in #8224
• remove the unused images train dialog and fix media page responsiveness by @mahiatlinux in #8215
• Studio: size FLUX.2-klein 9B LoRA training from its own weights, not the 4B family number by @danielhanchen in #8263
• Studio: LTX-2 video LoRA training from still images by @danielhanchen in #8196
• Say when attention has no fused kernel, and log the request behind a failed video by @danielhanchen in #8233
• Turn off embedding offload when the model is dispatched across devices by @danielhanchen in #8272
• Studio: keep the overlay stack off a maximised Live monitor by @danielhanchen in #8190
• Make the diffusion checkpoint-resume tests describe the host they run on by @danielhanchen in #8273
• Install the Linux CUDA sd-cli, and bump the pin two releases forward by @danielhanchen in #8208
• Studio: add MiniMax H3 video generation by @oobabooga in #7989
• Bump install.sh / install.ps1 pin to unsloth>=2026.8.10 by @danielhanchen in #8276
• Stop a GGUF pick downloading the dense base transformer it never opens by @danielhanchen in #8232
• Studio: let a model load be cancelled while it is still loading by @danielhanchen in #8243
• Studio: stop a multi-checkpoint GGUF repo collapsing into one row per quant by @danielhanchen in #8222
• Studio: refuse a diffusion or video load that cannot fit unified memory by @danielhanchen in #8213
• Studio: run MiniMax-H3's Qwen3-VL conditioner from the hosted INT8 checkpoint by @danielhanchen in #8283
• Fix image page pane and header alignment by @Imagineer99 in #8176
• Studio: match the loaded models card to the update banners by @shimmyshimmer in #8292
• Fix main: the two diffusers-import preflight tests only pass on a GPU host by @danielhanchen in #8281
• Desktop: fix audio files being rejected when dropped on the chat by @NilayYadav in #8271
• Fix two macOS installer bugs by @NilayYadav in #8279
• Studio: offer a per-canvas network grant when the offline CSP blocks a preview by @NilayYadav in #8262
• Read PSModulePath under the key Windows actually exports by @danielhanchen in #8304
• Studio: stop the sidebar flashing an icon rail it never collapses to by @danielhanchen in #8300
• Studio: keep the loaded models card in the corner it is supposed to be in by @danielhanchen in #8301
• Studio: say what the Train row is waiting for while it spins by @danielhanchen in #8302
• Fix image model dependency downloads and show full footprint by @Imagineer99 in #8036
• soften the dmg background glow and clear it off the icon label by @mahiatlinux in #8296
• Studio: remeasure the overlay stack when its content resizes in place by @danielhanchen in #8315
• Raise when SyntheticDataKit's vLLM server fails to start by @danielhanchen in #8297
• Only trust the adjacent ROCm-on-WSL helper for a real checkout run by @danielhanchen in #8316
• Studio: confirm a web_search call that fetches a url in auto mode by @danielhanchen in #8313
• Desktop: route native drops to the drop zone under the cursor by @NilayYadav in #8265
• Fix resume button missing after stop-and-save (#8150) by @Souravrajvi0 in #8217
• Studio: MiniMax-H3 follow-ups from the 7989 review by @danielhanchen in #8280
• Studio: run MiniMax-H3's INT8 denoiser from the ConvRot checkpoint by @danielhanchen in #8293
• Desktop: unify normal release updater flow by @wasimysaid in #8298
• Start the dmg halo taper under the icon so the disc still reads round by @danielhanchen in #8321
• Studio: show the files a tool call creates, and keep them in one place by @danielhanchen in #8256
• Studio: keep a rehearsal tool call quoted in markdown code as text by @danielhanchen in #8312
• CLI: verify-install reports a healthy install as broken around an editable checkout by @danielhanchen in #8308
• Studio: a base quant published at several bit widths is several variants by @danielhanchen in #8255
• Studio: make image-model companion assets reusable, visible, and safely removable by @danielhanchen in #8223
• Studio: let a dataset of clips be selected for training by @danielhanchen in #8287
• studio: offer the undistilled z-image base for lora training by @mahiatlinux in #8291
• Studio: better default Images and Video prompts by @shimmyshimmer in #8324
• CI: avoid blocking Windows releases on cache saves by @wasimysaid in #8325
• Studio: a binary request body should be a 422, not a 500 with the payload in the log by @danielhanchen in #8307
• Studio: stop two lines repeating in the server log for the whole session by @danielhanchen in #8306
• Studio: give the liveness polls one shared heartbeat instead of one each by @danielhanchen in #8310
• Studio: stop reporting the run's mean loss as the final step's loss by @danielhanchen in #8314
• studio: render a chat send without waiting on persistence by @mahiatlinux in #8136
• Studio: enable video generation on Apple Silicon by @Lyxot in #8198
• Studio: keep sidebar logo visible during training by @wasimysaid in #8331
• Absorb TRL config arguments retired between versions by @danielhanchen in #8330
• Studio: use Muse Glimmer's published sampling defaults by @danielhanchen in #8332
• Studio: keep third-party progress bars and raw trainer prints out of the structured log by @danielhanchen in #8311
• Video: make MiniMax-H3's Diffusers path fast by default by @danielhanchen in #8320
• Studio: give the MiniMax-H3 reference denoiser an int8 and fp8 rung by @danielhanchen in #8329

New Contributors

• @namdamdoi68-oss made their first contribution in #7891

Full Changelog: v0.1.52-beta...v0.1.60-beta