Auto compaction (preview) + LAN Remote Access

[danielhanchen]

Thanks for the support for Qwen3.8-27B and Unsloth Desktop last week! For this release, we merged 200+ PRs to introduce many new features, fixes including:

• Rolling Context + Auto Compaction (Experimental) for longer chats beyond context limits
• Remote & LAN Access (Preview) for easy network access without Cloudflare links
• Faster Chat - Improved streaming performance, reduced UI lag, and smoother long conversations.
• Support for custom llama.cpp builds. Toggles for Cache RAM, Mmap, Mlock, Checkpoints, Spe Decoding KV Cache, Vision On / Off
• Unsloth Dynamic v3.0 is released. New Qwen3.8-27B Dynamic v3.0 GGUFs deliver >10% higher top-1 accuracy compared to everyone else. Works with Unsloth.

Rolling context + auto compaction (Experimental)

You can now do long chats and surpass a model's context limit, with evicted turns
remaining searchable.

• Long local chats can continue past the context limit by rolling older turns out of
  active context.
• Auto compaction creates fresh context epochs instead of permanently trimming replies.
• Evicted conversations remain searchable, with improved fact recall.
• Context window information is shown before chat begins.

How it works

• Entire oldest turns are removed only when needed, never mid-message. The saved transcript stays unchanged.
• Evicted turns are indexed into a per-thread searchable archive using Unsloth's existing RAG pipeline (store, chunker, embedder, retrieval). Lexical search is prioritized because chat recall is usually exact matches like names, numbers, or IDs.
• A recall is forced during eviction instead of relying on the model to search itself. Later retrieval uses search_conversation.
• The archive persists across epochs, allowing future compactions to recover previously evicted context.
• No summarization: it showed little benefit and added ~190s per compaction.

Remote & LAN access (Preview)

Access Unsloth from another device on your network, managed from settings.

• New dedicated Settings section for remote access.
• Enable or disable LAN access without restarting.
• Supports connection addresses, QR codes, and optional auto-start.
• LAN access is disabled by default and requires changing the generated admin password.

Chat + Functionality improvements

• Long chats are faster with optimized streaming and thread handling.
• Projects organize chats, files, and workspaces together.
• Chats remember composer settings, system prompts, and model sampling options.
• Added prompt queueing and customizable keyboard shortcuts.
• Added edit_file tool for partial file modifications.
• Improved tool calling and MCP behavior.

Hardware + inference improvements

• Support for custom llama.cpp builds.
• Toggles for advanced settings - Cache RAM, Mmap, Mlock, Checkpoints, Speculative Decoding KV Cache, Vision On / Off
• Better validation for missing GPU support and oversized GGUF models.
• Improved MTP performance and VRAM handling.
• Better compatibility across ROCm, xFormers, and flash-attention setups.
• Added Intel XPU support.

API + serving

• Responses API supports structured output formatting.
• Improved recovery when llama-server crashes.
• Added OpenCode V2 support for unsloth start.

Hub + Desktop + training

• Hub supports more image/video workflows.
• Faster and improved GGUF downloads.
• Desktop packaging and permissions improved.
• Training scheduler and preprocessing improvements.

What's Changed

• Bump install.sh / install.ps1 pin to unsloth>=2026.8.18 by @danielhanchen in Bump install.sh / install.ps1 pin to unsloth>=2026.8.18 #8836
• Give the llama-flag fixtures the parallelSlotsClamped they now require by @vineethsaivs in Give the llama-flag fixtures the parallelSlotsClamped they now require #8837
• Studio: preserve pre-tool reasoning inside the GGUF tool loop by @losuler in Studio: preserve pre-tool reasoning inside the GGUF tool loop #8581
• docs: fix dead Linux .deb download link by @yzxcj797 in docs: fix dead Linux .deb download link #8891
• Model hub: show the Meta mark on the Unsloth re-upload of Muse Glimmer by @shimmyshimmer in Model hub: show the Meta mark on the Unsloth re-upload of Muse Glimmer #8949
• Studio: speed up the local model inventory and keep it off the API loop by @oobabooga in Studio: speed up the local model inventory and keep it off the API loop #8770
• Studio: reuse the cached GGUF work a load request has already done by @oobabooga in Studio: reuse the cached GGUF work a load request has already done #8771
• studio: keep each tool-call round separate when the delta index restarts by @mahiatlinux in studio: keep each tool-call round separate when the delta index restarts #8754
• studio: route bare tool-call fragments to the call that owns the index now by @mahiatlinux in studio: route bare tool-call fragments to the call that owns the index now #8755
• studio: opt-in model auto-switch for the image and video apis by @mahiatlinux in studio: opt-in model auto-switch for the image and video apis #8766
• Studio: keep project chats in Recents, add a Projects sidebar section by @shimmyshimmer in Studio: keep project chats in Recents, add a Projects sidebar section #8932
• Studio: find the real Documents folder on Windows for project workspaces by @shimmyshimmer in Studio: find the real Documents folder on Windows for project workspaces #8955
• fix(unsloth studio): only preprocess the rows a max_steps run will actually use instead of preprocessing validated datasets by @LeoBorcherding in fix(unsloth studio): only preprocess the rows a max_steps run will actually use instead of preprocessing validated datasets #8890
• Studio: stop the chat UI falling behind a fast stream by @oobabooga in Studio: stop the chat UI falling behind a fast stream #8845
• studio: keep each chat's composer pills and settings with the chat by @mahiatlinux in studio: keep each chat's composer pills and settings with the chat #8686
• studio: allow a per-connection max tokens limit on every provider by @mahiatlinux in studio: allow a per-connection max tokens limit on every provider #8944
• Crash guard: read dumpability in order, and stop losing crashes to scope and naming by @danielhanchen in Crash guard: read dumpability in order, and stop losing crashes to scope and naming #8808
• Studio: size the max_steps row bound for the data-parallel world by @danielhanchen in Studio: size the max_steps row bound for the data-parallel world #8958
• Studio: make the paste-to-attachment threshold a chat setting by @shimmyshimmer in Studio: make the paste-to-attachment threshold a chat setting #8963
• Fix CI on main: stale test doubles, a stale router stub, and two source defects by @danielhanchen in Fix CI on main: stale test doubles, a stale router stub, and two source defects #8956
• Make the startup profile a gate, with budgets from its own measurements by @danielhanchen in Make the startup profile a gate, with budgets from its own measurements #8965
• Keep xFormers working when flash-attn 4 is installed, and guard the varlen int32 overflow by @danielhanchen in Keep xFormers working when flash-attn 4 is installed, and guard the varlen int32 overflow #8957
• Stop six backend tests waiting on the clock instead of on a signal by @danielhanchen in Stop six backend tests waiting on the clock instead of on a signal #8975
• CI: stop three workflows paying for work they throw away by @danielhanchen in CI: stop three workflows paying for work they throw away #8976
• Give the pasted-text import an extension so the node test runner can load it by @danielhanchen in Give the pasted-text import an extension so the node test runner can load it #8981
• Make the packed-boundary guard reachable on the fused cross-entropy path by @danielhanchen in Make the packed-boundary guard reachable on the fused cross-entropy path #8959
• Studio: Fix embedded MTP performance under partial GPU offload by @oobabooga in Studio: Fix embedded MTP performance under partial GPU offload #8875
• Studio: tokenize the dataset online for plain-text single-pass runs by @danielhanchen in Studio: tokenize the dataset online for plain-text single-pass runs #8960
• Studio: keep K and V equal when the flash-attention retry resets the V cache by @danielhanchen in Studio: keep K and V equal when the flash-attention retry resets the V cache #8827
• Studio: hold the Auto-layers context exemption only while --fit runs by @danielhanchen in Studio: hold the Auto-layers context exemption only while --fit runs #8825
• Studio: stop a sent prompt staying in the chat composer by @shimmyshimmer in Studio: stop a sent prompt staying in the chat composer #8849
• Chat: queue prompts with Cmd/Ctrl+Enter and drag to reorder the queue by @shimmyshimmer in Chat: queue prompts with Cmd/Ctrl+Enter and drag to reorder the queue #8952
• Studio: repin the permission-mode contract to behaviour, not a declaration by @danielhanchen in Studio: repin the permission-mode contract to behaviour, not a declaration #8986
• Cut Repo tests (CPU) roughly in half: stop the synthetic vLLM tests reaping a fake server by @danielhanchen in Cut Repo tests (CPU) roughly in half: stop the synthetic vLLM tests reaping a fake server #8993
• Core: sweep every transformers model_type across a process pool by @danielhanchen in Core: sweep every transformers model_type across a process pool #8995
• Repair the Tauri retry test against the download-marker pipeline by @danielhanchen in Repair the Tauri retry test against the download-marker pipeline #8996
• Studio: read every launcher world size when resolving a step-capped run's passes by @danielhanchen in Studio: read every launcher world size when resolving a step-capped run's passes #9000
• Repair the prompt-queue contract test against the queueing refactor by @danielhanchen in Repair the prompt-queue contract test against the queueing refactor #9026
• studio: stop prompting to stop chats for a model already loaded by @mahiatlinux in studio: stop prompting to stop chats for a model already loaded #8943
• Studio: remember chat parameters per model by @shimmyshimmer in Studio: remember chat parameters per model #8757
• Studio: share attached files across a project by @shimmyshimmer in Studio: share attached files across a project #8756
• Fix a 500 on dictation Unload before any backend is resident by @danielhanchen in Fix a 500 on dictation Unload before any backend is resident #9013
• Run the repo CPU test suite on all four runner cores by @danielhanchen in Run the repo CPU test suite on all four runner cores #9019
• Scan package archives across cores instead of one at a time by @danielhanchen in Scan package archives across cores instead of one at a time #9024
• Stop TestProbeGating asserting a tier the mapping check is meant to upgrade by @danielhanchen in Stop TestProbeGating asserting a tier the mapping check is meant to upgrade #9029
• Stub the implicit STT load in the disconnect-cancellation test by @danielhanchen in Stub the implicit STT load in the disconnect-cancellation test #9031
• Stop test_begin_load_never_refuses_auto racing the thread it is asserting about by @danielhanchen in Stop test_begin_load_never_refuses_auto racing the thread it is asserting about #9034
• Measure the streaming stripper against CPU time, not wall clock by @danielhanchen in Measure the streaming stripper against CPU time, not wall clock #9035
• Fix silent lr_warmup_steps no-op under the default constant scheduler by @Eoin-Houstoun in Fix silent lr_warmup_steps no-op under the default constant scheduler #8464
• Repair the contract tests studio: stop prompting to stop chats for a model already loaded #8943 left behind on main by @danielhanchen in Repair the contract tests #8943 left behind on main #9067
• Take the loaded-models indicator suite off the UI job's critical path by @danielhanchen in Take the loaded-models indicator suite off the UI job's critical path #9060
• Keep the killed-by-signal exit status when release() fails by @danielhanchen in Keep the killed-by-signal exit status when release() fails #9072
• Put the event-loop latency bound between the two things it separates by @danielhanchen in Put the event-loop latency bound between the two things it separates #9077
• Stub unsloth in test_safetensors_reasoning_stream so it can run first by @danielhanchen in Stub unsloth in test_safetensors_reasoning_stream so it can run first #9027
• Report what the launcher logged when a signal test fails by @danielhanchen in Report what the launcher logged when a signal test fails #9079
• Studio: let Allow microphone recover from a saved "Don't allow" on Windows by @shimmyshimmer in Studio: let Allow microphone recover from a saved "Don't allow" on Windows #9006
• Run both ends of the interpreter matrix on a pull request, all four on main by @danielhanchen in Run both ends of the interpreter matrix on a pull request, all four on main #9080
• fix(studio): enable microphone/camera access in Linux WebKitGTK webview by @rsd-darshan in fix(studio): enable microphone/camera access in Linux WebKitGTK webview #8720
• Studio: budget the JavaScript that runs before the first screen by @danielhanchen in Studio: budget the JavaScript that runs before the first screen #8964
• fix(studio): derive disk percentage from displayed usage by @dyuan311 in fix(studio): derive disk percentage from displayed usage #9021
• Studio: stop rereading the whole reply on every streamed arrival by @danielhanchen in Studio: stop rereading the whole reply on every streamed arrival #9012
• Studio: stop every message part re-rendering on each streaming chunk by @danielhanchen in Studio: stop every message part re-rendering on each streaming chunk #9014
• Studio: stop a LaTeX rewrite throwing away the whole retained markdown prefix by @danielhanchen in Studio: stop a LaTeX rewrite throwing away the whole retained markdown prefix #9017
• Studio: stop later code spans changing whether an earlier one is escaped by @danielhanchen in Studio: stop later code spans changing whether an earlier one is escaped #9028
• Chat: stop a message delete from re-rendering the whole thread by @danielhanchen in Chat: stop a message delete from re-rendering the whole thread #9042
• Chat: stop a keystroke in the composer costing one pass over the thread by @danielhanchen in Chat: stop a keystroke in the composer costing one pass over the thread #9054
• Let every main run finish instead of cancelling it before it starts by @danielhanchen in Let every main run finish instead of cancelling it before it starts #9082
• One multi-turn smoke script for all three operating systems by @danielhanchen in One multi-turn smoke script for all three operating systems #9086
• Studio: give a CRLF reply the same incremental rendering as an LF one by @danielhanchen in Studio: give a CRLF reply the same incremental rendering as an LF one #9088
• Studio: sweep the streaming markdown corpus at every prefix by @danielhanchen in Studio: sweep the streaming markdown corpus at every prefix #9085
• Do not let the signal handler's own logging stop it dying by @danielhanchen in Do not let the signal handler's own logging stop it dying #9083
• Stop the media switch test racing a 0.3s budget against 2s of cold path by @danielhanchen in Stop the media switch test racing a 0.3s budget against 2s of cold path #9097
• Studio: honor the Responses API text.format structured output by @NilayYadav in Studio: honor the Responses API text.format structured output #8846
• Studio: stop the streaming prefix checks scanning the whole reply by @danielhanchen in Studio: stop the streaming prefix checks scanning the whole reply #9038
• Studio: stop the streamed reply being flattened on every arrival by @danielhanchen in Studio: stop the streamed reply being flattened on every arrival #9049
• Studio: strip the trailing template placeholder once, on the finished reply by @danielhanchen in Studio: strip the trailing template placeholder once, on the finished reply #9101
• Studio: fix the Windows-only path failures in the frontend test suite, and catch the shape from Linux by @danielhanchen in Studio: fix the Windows-only path failures in the frontend test suite, and catch the shape from Linux #8980
• Studio: run the frontend unit tests on Windows in CI by @danielhanchen in Studio: run the frontend unit tests on Windows in CI #9099
• Run the Backend CI matrix in parallel, like its sibling job already does by @danielhanchen in Run the Backend CI matrix in parallel, like its sibling job already does #9095
• Pair the diffusion warmup presets with a scheduler that uses them by @vineethsaivs in Pair the diffusion warmup presets with a scheduler that uses them #8593
• Make the models a defaults file claims actually load it by @vineethsaivs in Make the models a defaults file claims actually load it #8726
• Derive warmup steps from a model default that gives a ratio by @vineethsaivs in Derive warmup steps from a model default that gives a ratio #8723
• Studio: let the hub run non-GGUF image and video models by @NilayYadav in Studio: let the hub run non-GGUF image and video models #8855
• Restore crypto.randomUUID before the Studio bundle evaluates by @Lyxot in Restore crypto.randomUUID before the Studio bundle evaluates #9075
• studio: refuse a gguf that cannot fit in free vram plus available ram by @mahiatlinux in studio: refuse a gguf that cannot fit in free vram plus available ram #8883
• studio: add settings-managed LAN access by @mahiatlinux in studio: add settings-managed LAN access #8951
• Ask the repo, not sys.path, whether a stubbed name is third-party by @danielhanchen in Ask the repo, not sys.path, whether a stubbed name is third-party #9133
• Stop the sidecar tests writing into the real Studio install by @danielhanchen in Stop the sidecar tests writing into the real Studio install #9131
• Delete a test that never ran, and stop another appearing by @danielhanchen in Delete a test that never ran, and stop another appearing #9135
• Make a hung kaggle launcher say where it is stuck by @danielhanchen in Make a hung kaggle launcher say where it is stuck #9104
• Stop the banner layout suite waiting out the update timer on every boot by @danielhanchen in Stop the banner layout suite waiting out the update timer on every boot #9105
• Give each xdist worker its own torch.compile cache directory by @danielhanchen in Give each xdist worker its own torch.compile cache directory #9134
• One interpreter leg on a pull request, and a floor lint that reads more than syntax by @danielhanchen in One interpreter leg on a pull request, and a floor lint that reads more than syntax #9100
• Split Chat UI Tests into four shards on the Studio boundaries by @danielhanchen in Split Chat UI Tests into four shards on the Studio boundaries #9132
• Bundle Unsloth GGUF CI onto one runner, matching Windows and macOS by @danielhanchen in Bundle Unsloth GGUF CI onto one runner, matching Windows and macOS #9139
• Give the small Windows checks one box per image instead of one box per check by @danielhanchen in Give the small Windows checks one box per image instead of one box per check #9143
• Let tests skip the settle wait between retried VRAM reads by @danielhanchen in Let tests skip the settle wait between retried VRAM reads #9141
• Studio: keep the streaming render harness the perf PRs kept rebuilding by @danielhanchen in Studio: keep the streaming render harness the perf PRs kept rebuilding #8969
• Studio: pin why a Streamdown remount keeps its highlighted code by @danielhanchen in Studio: pin why a Streamdown remount keeps its highlighted code #9048
• Chat: cover which messages a delete takes with it by @danielhanchen in Chat: cover which messages a delete takes with it #9087
• Studio: pin what the trailing-placeholder window does with nested placeholders by @danielhanchen in Studio: pin what the trailing-placeholder window does with nested placeholders #9091
• Studio: stop a settings module reading a chat store key before it exists by @danielhanchen in Studio: stop a settings module reading a chat store key before it exists #8979
• Studio: load the settings tab panels when they are shown, not at launch by @danielhanchen in Studio: load the settings tab panels when they are shown, not at launch #8966
• Name the encoding when reading wslinfo's output by @danielhanchen in Name the encoding when reading wslinfo's output #9142
• Studio: keep pandas out of the backend startup import graph by @danielhanchen in Studio: keep pandas out of the backend startup import graph #8962
• Run the three loaded-models-indicator engines at once by @danielhanchen in Run the three loaded-models-indicator engines at once #9158
• Unblock Mac Studio GGUF CI, red on every main run since studio: refuse a gguf that cannot fit in free vram plus available ram #8883 by @danielhanchen in Unblock Mac Studio GGUF CI, red on every main run since #8883 #9155
• Drop the Exa preset from the MCP menu by @shimmyshimmer in Drop the Exa preset from the MCP menu #9160
• Studio: stop the overlay rail going click-through while it is scrolling by @danielhanchen in Studio: stop the overlay rail going click-through while it is scrolling #9150
• Open the connection form when Settings > Connections has nothing to list by @shimmyshimmer in Open the connection form when Settings > Connections has nothing to list #9157
• Security audit: pin openai, and re-review the four digest-pinned entries by @danielhanchen in Security audit: pin openai, and re-review the four digest-pinned entries #9148
• Say why a Xet download sits at 0% until it finishes by @shimmyshimmer in Say why a Xet download sits at 0% until it finishes #9159
• Stop running every macOS workflow on every commit to main by @danielhanchen in Stop running every macOS workflow on every commit to main #9174
• Time the install from CI, without changing the installers by @danielhanchen in Time the install from CI, without changing the installers #9153
• Studio: measure where a heavy thread stalls, across engines and thread size by @danielhanchen in Studio: measure where a heavy thread stalls, across engines and thread size #9016
• Studio CI: make a failing browser smoke say why, and stop it skipping the build gates by @danielhanchen in Studio CI: make a failing browser smoke say why, and stop it skipping the build gates #8983
• Stop spending the Actions cache budget on caches nothing reads by @danielhanchen in Stop spending the Actions cache budget on caches nothing reads #9151
• Fix low VRAM mmproj fallback in Studio by @wasimysaid in Fix low VRAM mmproj fallback in Studio #9173
• Run two short jobs as background lanes of Lint CI, not on their own runners by @danielhanchen in Run two short jobs as background lanes of Lint CI, not on their own runners #9176
• Studio: fail fast when the installed llama.cpp prebuilt has no kernels for this GPU by @NilayYadav in Studio: fail fast when the installed llama.cpp prebuilt has no kernels for this GPU #8841
• Studio: bulk chat management multi-select move/pin/archive/export/delete by @NilayYadav in Studio: bulk chat management multi-select move/pin/archive/export/delete #8745
• hide app-only MCP tools from the model and explain why local commands are disabled by @NilayYadav in hide app-only MCP tools from the model and explain why local commands are disabled #8739
• Studio: remember last used model in the backend, not per-browser by @NilayYadav in Studio: remember last used model in the backend, not per-browser #8587
• Studio: show MCP server display names instead of internal ids by @NilayYadav in Studio: show MCP server display names instead of internal ids #8557
• Studio: make the chat thread stop getting slower as it fills by @danielhanchen in Studio: make the chat thread stop getting slower as it fills #8992
• Studio: recover marked reports across reasoning channels by @alkinun in Studio: recover marked reports across reasoning channels #9119
• show the local model's context window before the first token count by @mahiatlinux in show the local model's context window before the first token count #8882
• studio: stop nudging a turn that asks the user for a missing detail by @mahiatlinux in studio: stop nudging a turn that asks the user for a missing detail #8946
• Say both load fallbacks when both fire, and repair the two suites that guard it by @danielhanchen in Say both load fallbacks when both fire, and repair the two suites that guard it #9189
• Run the Mac GGUF inference phases in the Mac UI job, not on a second runner by @danielhanchen in Run the Mac GGUF inference phases in the Mac UI job, not on a second runner #9213
• Print what the server said when an inference smoke request 4xx's by @danielhanchen in Print what the server said when an inference smoke request 4xx's #9202
• Allowlist huggingface-hub's http_backoff loop so Security audit stops failing by @danielhanchen in Allowlist huggingface-hub's http_backoff loop so Security audit stops failing #9252
• Count a capped overlay stack by what it reads, not by how the cap is spelled by @danielhanchen in Count a capped overlay stack by what it reads, not by how the cap is spelled #9253
• Studio: support llama.cpp semantic version output by @oobabooga in Studio: support llama.cpp semantic version output #9127
• Bound every apt step in CI, and fix the retry that never ran by @danielhanchen in Bound every apt step in CI, and fix the retry that never ran #9256
• Make apt fail fast, so the retry has something to retry by @danielhanchen in Make apt fail fast, so the retry has something to retry #9260
• Record the venv own requirement digests, not the installers by @danielhanchen in Record the venv own requirement digests, not the installers #9263
• Stub all three fields the snapshot reads, not two of them by @danielhanchen in Stub all three fields the snapshot reads, not two of them #9274
• Ask whether Playwright system libraries are missing before installing them by @danielhanchen in Ask whether Playwright system libraries are missing before installing them #9273
• Stop blaming the recipe for a turn that never came back by @danielhanchen in Stop blaming the recipe for a turn that never came back #9266
• Move the two version-incidental macOS jobs onto the emptier pool by @danielhanchen in Move the two version-incidental macOS jobs onto the emptier pool #9277
• Give the Colab freeze a pip cache, like the job beside it already has by @danielhanchen in Give the Colab freeze a pip cache, like the job beside it already has #9276
• Wait for the Compare nav instead of sampling for it by @danielhanchen in Wait for the Compare nav instead of sampling for it #9264
• Install the browser engines each Chat UI shard actually drives by @danielhanchen in Install the browser engines each Chat UI shard actually drives #9283
• Update README to include Intel XPU support description for unsloth studio by @JoshuaL3000 in Update README to include Intel XPU support description for unsloth studio #9250
• Allow unlimited Deep Research time budgets by @alkinun in Allow unlimited Deep Research time budgets #9078
• Stop the orphan-scan tests asking the host whether an invented PID is alive by @danielhanchen in Stop the orphan-scan tests asking the host whether an invented PID is alive #9280
• Repair the tests main is currently red on by @danielhanchen in Repair the tests main is currently red on #9192
• Re-type the native default, not whatever the input happens to show by @danielhanchen in Re-type the native default, not whatever the input happens to show #9282
• Studio: confirm an Audio model swap instead of dead-ending on the backend 409 by @shimmyshimmer in Studio: confirm an Audio model swap instead of dead-ending on the backend 409 #9092
• fix(studio): refuse a hand-set context unified memory cannot hold by @danielhanchen in fix(studio): refuse a hand-set context unified memory cannot hold #9172
• Studio: size Xet download buffers from free RAM, not total by @shimmyshimmer in Studio: size Xet download buffers from free RAM, not total #9052
• Studio: make the partial download affordance tell the truth by @shimmyshimmer in Studio: make the partial download affordance tell the truth #8982
• Cache uv's downloads, which is now the largest single cost in CI by @danielhanchen in Cache uv's downloads, which is now the largest single cost in CI #9271
• Studio: incrementally tokenize streaming code fences by @oobabooga in Studio: incrementally tokenize streaming code fences #8935
• Desktop: ship a complete Linux AppImage by @oobabooga in Desktop: ship a complete Linux AppImage #9113
• Studio: link the Read aloud TTS row to Audio and show the preview generate wait by @shimmyshimmer in Studio: link the Read aloud TTS row to Audio and show the preview generate wait #9093
• studio: show the model a chat was started on, and offer it back by @shimmyshimmer in studio: show the model a chat was started on, and offer it back #9081
• Desktop: make every drop zone take a drop again ([Bug] Unsloth Desktop: Drag and Drop doesn't work all the time with JPG / PNG etc.. #9036) by @shimmyshimmer in Desktop: make every drop zone take a drop again (#9036) #9056
• Studio: say when a scan folder cannot be read instead of showing no models by @shimmyshimmer in Studio: say when a scan folder cannot be read instead of showing no models #9053
• Studio: recover the OpenAI passthrough from a crashed llama-server by @shimmyshimmer in Studio: recover the OpenAI passthrough from a crashed llama-server #8999
• Studio: stop the desktop health watchdog killing a backend that is busy generating by @shimmyshimmer in Studio: stop the desktop health watchdog killing a backend that is busy generating #8990
• Studio: add a keyboard shortcuts page and make the existing chords rebindable by @shimmyshimmer in Studio: add a keyboard shortcuts page and make the existing chords rebindable #8948
• Scroll into view before a forced click by @danielhanchen in Scroll into view before a forced click #9291
• studio: keep each chat's sampling params and system prompt with the chat by @shimmyshimmer in studio: keep each chat's sampling params and system prompt with the chat #9055
• Run the zoo suite in parallel, minus the two files that cannot share a worker by @danielhanchen in Run the zoo suite in parallel, minus the two files that cannot share a worker #9285
• Cache the .deb set webkit needs, so a dead mirror costs a miss not a shard by @danielhanchen in Cache the .deb set webkit needs, so a dead mirror costs a miss not a shard #9289
• Revert "Say why a Xet download sits at 0% until it finishes" by @shimmyshimmer in Revert "Say why a Xet download sits at 0% until it finishes" #9293
• Default preserve thinking on for Qwen3.8 by @Imagineer99 in Default preserve thinking on for Qwen3.8 #9096
• Studio: price a partial GGUF by what is left to fetch by @shimmyshimmer in Studio: price a partial GGUF by what is left to fetch #8989
• Studio: add an edit_file tool so agents stop rewriting whole files by @shimmyshimmer in Studio: add an edit_file tool so agents stop rewriting whole files #8753
• fix(studio): img2img/inpaint on GGUF image models — from_pipe fallback with no recast (Studio 0.1.800-beta: img2img/inpaint fails with 'Casting a quantized model to a new dtype is unsupported' for GGUF image models #9186) by @yzxcj797 in fix(studio): img2img/inpaint on GGUF image models — from_pipe fallback with no recast (#9186) #9193
• Install Studio on a Windows host whose C# compiler cannot run by @danielhanchen in Install Studio on a Windows host whose C# compiler cannot run #9178
• fix(AMD): retry ROCm llama-server with bundled HIP on HIP/ROCR mismatch by @indrajeetapache in fix(AMD): retry ROCm llama-server with bundled HIP on HIP/ROCR mismatch #9002
• Studio: use a custom llama.cpp build by @Etherll in Studio: use a custom llama.cpp build #9292
• Studio: keep health checks responsive during Anthropic generation by @Imagineer99 in Studio: keep health checks responsive during Anthropic generation #9201
• Studio: propagate required backend version to repair pipeline ([Bug] Error on second launch of macos app #8610) by @karan68 in Studio: propagate required backend version to repair pipeline (#8610) #8670
• Keep a sampler's std and uuid format across a save and reopen by @vineethsaivs in Keep a sampler's std and uuid format across a save and reopen #9225
• Studio: never pick a macOS AppleDouble sidecar as a GGUF by @sts-change in Studio: never pick a macOS AppleDouble sidecar as a GGUF #8919
• Studio: follow up Studio: keep health checks responsive during Anthropic generation #9201 across non-streaming sibling paths by @oobabooga in Studio: follow up #9201 across non-streaming sibling paths #9312
• Studio: take common SQLite reads off the event loop thread by @oobabooga in Studio: take common SQLite reads off the event loop thread #9234
• Keep xFormers attention masks on the GPU running each layer by @oobabooga in Keep xFormers attention masks on the GPU running each layer #8516
• Studio: add rolling context windows for local GGUF chat by @alkinun in Studio: add rolling context windows for local GGUF chat #8961
• Studio: keep and search the turns rolling context evicts by @danielhanchen in Studio: keep and search the turns rolling context evicts #9074
• Studio: recall the latest version of a fact, not the most quotable one by @danielhanchen in Studio: recall the latest version of a fact, not the most quotable one #9161
• Studio: compact a chat by resetting the epoch, not by trimming it forever by @danielhanchen in Studio: compact a chat by resetting the epoch, not by trimming it forever #9162
• Support OpenCode V2 in unsloth start by @pascalandr in Support OpenCode V2 in unsloth start #9275
• Keep path_utils off PEP 604, which the 3.9 floor gate rejects by @danielhanchen in Keep path_utils off PEP 604, which the 3.9 floor gate rejects #9335
• Stub the two helpers the sliced harnesses started importing by @danielhanchen in Stub the two helpers the sliced harnesses started importing #9342
• Put back the AppleDouble GGUF filters Studio: keep and search the turns rolling context evicts #9074 reverted, and guard them by @danielhanchen in Put back the AppleDouble GGUF filters #9074 reverted, and guard them #9353
• Fix the two backend guards that Studio: use a custom llama.cpp build #9292 turned red on the 3.13 leg by @danielhanchen in Fix the two backend guards that #9292 turned red on the 3.13 leg #9348
• Drain the sampling debounce long enough for the node CI actually runs by @danielhanchen in Drain the sampling debounce long enough for the node CI actually runs #9332
• Take two ubuntu jobs off their own runners: absorb one, delete the other by @danielhanchen in Take two ubuntu jobs off their own runners: absorb one, delete the other #9360
• Studio: stop docked composer from squeezing download panel by @oobabooga in Studio: stop docked composer from squeezing download panel #9334
• Drain the sampling suites on the store's write chains, not on node's loader by @danielhanchen in Drain the sampling suites on the store's write chains, not on node's loader #9367
• Say which runtime components an AppImage is missing before its env hygiene by @danielhanchen in Say which runtime components an AppImage is missing before its env hygiene #9370
• Drop the HunyuanImage GGUF the Hub no longer serves by @danielhanchen in Drop the HunyuanImage GGUF the Hub no longer serves #9369
• Studio: fix two intermittently failing tests by @danielhanchen in Studio: fix two intermittently failing tests #9200
• Acknowledge the Colab rotation the oracle snapshot is measuring by @danielhanchen in Acknowledge the Colab rotation the oracle snapshot is measuring #9376
• Studio: carry a restated instruction once, not eight times by @danielhanchen in Studio: carry a restated instruction once, not eight times #9368
• Stop 13 jobs rebuilding the same frontend on every commit by @danielhanchen in Stop 13 jobs rebuilding the same frontend on every commit #9375
• Stop 19 test files racing on one PowerShell startup cache by @danielhanchen in Stop 19 test files racing on one PowerShell startup cache #9371
• Stop the pip cache key duplicating itself on every Python patch bump by @danielhanchen in Stop the pip cache key duplicating itself on every Python patch bump #9377
• Studio: list Ollama models in the chat picker and load their manifest refs by @lonexreb in Studio: list Ollama models in the chat picker and load their manifest refs #9237
• Stop the Studio temp probe leaving anything behind by @danielhanchen in Stop the Studio temp probe leaving anything behind #9298
• Stop the dist cache key hashing 457 files the rebuild check never reads by @danielhanchen in Stop the dist cache key hashing 457 files the rebuild check never reads #9380
• Studio: overturn a chat-only MLX verdict the stack contradicts ([Bug] Studio: MLX Train/Export falsely greyed out — startup thread race on first transformers import (not a broken install) #9120) by @yzxcj797 in Studio: overturn a chat-only MLX verdict the stack contradicts (#9120) #9124
• Put the managed Node bin dir on PATH when spawning stdio MCP servers by @NilayYadav in Put the managed Node bin dir on PATH when spawning stdio MCP servers #9304
• Studio: normalize replayed tool-call ids for external providers by @NilayYadav in Studio: normalize replayed tool-call ids for external providers #9116
• Don't select Flash Attention 2 for float32 model loads by @NilayYadav in Don't select Flash Attention 2 for float32 model loads #9110
• Studio: keep every cached revision's quants on the on-device card by @Lyxot in Studio: keep every cached revision's quants on the on-device card #9381
• Studio: put Remote and LAN access on their own settings tab by @shimmyshimmer in Studio: put Remote and LAN access on their own settings tab #9389
• Run six pinned-symbol jobs on one runner instead of six by @danielhanchen in Run six pinned-symbol jobs on one runner instead of six #9385
• Studio: take the model name back off the chat rows by @shimmyshimmer in Studio: take the model name back off the chat rows #9395
• Stop five Windows jobs rebuilding the same frontend on every commit by @danielhanchen in Stop five Windows jobs rebuilding the same frontend on every commit #9387
• Run the Windows UI suites in two parallel lanes instead of five in sequence by @danielhanchen in Run the Windows UI suites in two parallel lanes instead of five in sequence #9391
• Studio: keep a downloaded model listed when upstream moves on by @oobabooga in Studio: keep a downloaded model listed when upstream moves on #9347
• Studio: offer only the Codex models the ChatGPT plan can actually reach by @NilayYadav in Studio: offer only the Codex models the ChatGPT plan can actually reach #9302
• Studio: place the vision projector on the CPU only when it does not fit in VRAM by @danielhanchen in Studio: place the vision projector on the CPU only when it does not fit in VRAM #9063
• Fix Qwen3.8 presence penalty defaults by @Imagineer99 in Fix Qwen3.8 presence penalty defaults #9372
• Studio: Forward thinking controls and return reasoning on /v1/messages and /v1/responses by @NilayYadav in Studio: Forward thinking controls and return reasoning on /v1/messages and /v1/responses #9107
• Studio: keep the working llama.cpp install when a prebuilt update can't move it aside by @NilayYadav in Studio: keep the working llama.cpp install when a prebuilt update can't move it aside #9221
• Studio: add a Disable Vision toggle to a model's Advanced Settings by @danielhanchen in Studio: add a Disable Vision toggle to a model's Advanced Settings #9383
• Studio: stop the scroll stabilizer invalidating style for the whole thread by @danielhanchen in Studio: stop the scroll stabilizer invalidating style for the whole thread #9328
• Studio: never offer a calibration imatrix as a GGUF variant by @danielhanchen in Studio: never offer a calibration imatrix as a GGUF variant #9409
• Stop Codex cancelling its own first turn while llama-server reads the prompt by @danielhanchen in Stop Codex cancelling its own first turn while llama-server reads the prompt #9403
• Studio: stop a reset dropping the task when the user types short prompts by @danielhanchen in Studio: stop a reset dropping the task when the user types short prompts #9379
• Studio: finish a reply that hit Max Tokens instead of asking whether to by @danielhanchen in Studio: finish a reply that hit Max Tokens instead of asking whether to #9382
• Format the mmproj placement policy tests so pre-commit stops failing every PR by @danielhanchen in Format the mmproj placement policy tests so pre-commit stops failing every PR #9414
• Studio: scope the sidebar drag's custom-property writes off the thread's ancestors by @danielhanchen in Studio: scope the sidebar drag's custom-property writes off the thread's ancestors #9400
• Studio: size a large tool result to the window the model is running with by @danielhanchen in Studio: size a large tool result to the window the model is running with #9384
• Report what the model server actually said when a stream fails by @danielhanchen in Report what the model server actually said when a stream fails #9390
• Admit generations against the KV cache, not just the slot count by @danielhanchen in Admit generations against the KV cache, not just the slot count #9392
• Studio: add Mmap/Mlock, draft KV cache dtype, Checkpoints and Cache RAM to Run settings by @danielhanchen in Studio: add Mmap/Mlock, draft KV cache dtype, Checkpoints and Cache RAM to Run settings #9410

New Contributors

• @losuler made their first contribution in Studio: preserve pre-tool reasoning inside the GGUF tool loop #8581
• @yzxcj797 made their first contribution in docs: fix dead Linux .deb download link #8891
• @Eoin-Houstoun made their first contribution in Fix silent lr_warmup_steps no-op under the default constant scheduler #8464
• @dyuan311 made their first contribution in fix(studio): derive disk percentage from displayed usage #9021
• @indrajeetapache made their first contribution in fix(AMD): retry ROCm llama-server with bundled HIP on HIP/ROCR mismatch #9002
• @karan68 made their first contribution in Studio: propagate required backend version to repair pipeline (#8610) #8670
• @sts-change made their first contribution in Studio: never pick a macOS AppleDouble sidecar as a GGUF #8919
• @pascalandr made their first contribution in Support OpenCode V2 in unsloth start #9275
• @lonexreb made their first contribution in Studio: list Ollama models in the chat picker and load their manifest refs #9237

Full Changelog: v0.1.800-beta...v0.1.801-beta

---

This discussion was created from the release Auto compaction (preview) + LAN Remote Access.

[konradbr]

Hi @danielhanchen thank you for the update & especially the rolling context window work. I've added a feature request in this regard with as useful a specification as I could put together right now here: #10345

You've made a few assumptions & decisions, e.g.:

• Evicted turns are indexed into a per-thread searchable archive using Unsloth's existing RAG pipeline (store, chunker, embedder, retrieval).
• Lexical search is prioritized because chat recall is usually exact matches like names, numbers, or IDs.
• No summarization: it showed little benefit and added ~190s per compaction.

In my case (and probably in many others') the assumption that chat recall is usually exact matches doesn't hold true. And often to be able to search for a specific match, the model has to remember in the first place that such topic was part of the thread.

I'm not sure what data you tested on, but I believe that in many cases (including mine), short summarizations would add a lot(!!) of value. Did you mean to write 190s or 190ms? In my experience, the compaction takes comparitively little time.

All the best, K.Br.