LU v2.6.0

This is an old release. Get the current version from the latest release, or download the Windows installer straight from lu-labs.ai. Older builds miss fixes and features, and some of them predate the Cloud.

🎉 lu 2.6.0 is out

221 commits since 2.5.10. most of it came out of running the shipped build end to end on a real 12 gb card and working through the reports from discord and github. auto update picks this up on your next launch.

video

gguf video models generate for the first time. the catalog offered gguf video bundles at 19 gb a piece, and the video pipeline only read the two comfyui loaders that list safetensors files, so an installed gguf quant could never render a clip. the gguf loaders are wired in now, the video lane shows its starter card on an empty model list instead of a blank pane, and a freshly booted app no longer races its own model list when a persisted pick loads slower than the first click.

video renders finish on 12 gb cards. after sampling, the vae decode of a video ran full frame next to the resident unet, got paged by the windows driver instead of a clean out of memory, and sat at 100% gpu for an hour without finishing. every video decode now runs tiled whenever the installed comfyui has the tiled node, which turns that hour into minutes. image decodes are untouched.

the render progress bar ticks. progress only repainted on comfyui events, and the long stretches like model loading and frame decoding send none for minutes, so the bar froze on one label and looked hung while the gpu was at 100%. a one second ticker keeps the phase label and the elapsed time moving, and gguf loads show their loading phase too.

models and catalog

deepseek v4 flash, in both lanes. in lu cloud for pro and max, served through the same server driven catalog as the rest, so it reaches installed apps without an update. locally the catalog now carries the official 0731 release in three sizes next to the abliterated build that was already there. it is a 284b moe, so the local quants start at 82 gb and want a big rig; the qwen 3.5 9b distill in the catalog is the small way in.

a multi part gguf is one model, not several broken ones. sharded downloads install and load as a single model, and the sharded download flow is covered by a ui test now.

a hugging face token rides along on every hub download, and an aborted mlx download stops counting as installed.

chat, agent and code

agent and code run on the unrestricted models. they were hidden on exactly the models people come here for. the cloud catalog states tool support per model now, and code and agent stream on every transport, hermes included.

consecutive tool calls render as one morphing band instead of a wall of boxes, and the picker says when a model context window is too small for a tool set instead of failing on send.

regenerate replaces the answer instead of asking the same question again, deleting a chat line is findable (hover or right click), and a dead built in engine says what is wrong and which model is missing instead of leaking a proxy error.

mcp

external mcp servers could not work in any shipped build. three layers: the module error from the reports, a bundler directive that only held in dev builds, and a permission list that allowed no launcher to spawn at all. all three are closed, so a server you add in settings reaches the agent tool list.

mac

the local lane on macos: mlx models show up in the model manager instead of hidden rails, create can set up local media itself, local renders survive a restart instead of turning into dead tiles, download on a local render saves the file, the chat agent sees the image it just generated, quitting stops the child processes, and the bundle carries a real ad hoc signature so permission answers stick.

the built in engine

an expert panel for context size, flash attention, kv quant, threads, gpu layers, mlock and mmap. an honest token counter with a context dropdown. context presets capped at what the gguf header allows. ollama models are evicted before a start so the vram is free, comfyui vram is released too, and the engine refuses to adopt a stranger sitting on its port. picking a different gguf on the models page swaps the engine.

downloads

big downloads no longer die in a fixed timeout, and an interrupted download resumes where it stopped. two starts on one file name can no longer corrupt a model. install progress shows in the download tray at full width with clean sizes.

voice

read aloud installs its piper voice while comfyui is running and names the cause when it fails. a mic below 16 khz no longer writes a wav header that lies. a timed out dictation stops handing its text to the next one. read aloud stops cutting chinese and japanese mid word.

security

a percent encoded path can no longer walk past the remote permission gate, and a glob pattern gets the same jail as a plain path. both have tests now. keychain calls moved off the main thread, so a blocked vault no longer freezes the app.

the rest

document chat handles documents without punctuation. vram above 4 gb is no longer invented. the benchmark table can be cleared, exported and finished from the ui. the agent got a screenshot tool. character studio trains and generates on local hardware. the linux appimage no longer leaks its bundled libraries into every python it starts, which was breaking comfyui installs for appimage users. a workflow condition step no longer always takes the else branch. a background agent task runs in the project folder instead of the lu folder. the wallet runs on your billing period instead of the calendar month.

web studio

qwen image edit in the cloud edit lane: describe the change and it edits the whole frame, no mask needed. the masked flux inpaint stays.

hopefully this fixes it for you, and if something is still broken, we are in the discord and we answer.