Strata 0.2.15 arrives as a bug-fix release for the engine running 177B-parameter models on cheap GPUs
Strata 0.2.15 is live on GitHub now, and for current users the news is almost too boring to report. Update as you normally would, and your models, chats, and settings carry over untouched.
Another Strata update has been released today, following yesterday's Strata 0.2.14 release. Strata positions itself as an inference engine plus a desktop app and a terminal tool, pitched simply as a way to run powerful AI locally. No account. No cloud. Everything runs on localhost by default, and it exposes OpenAI-compatible and Anthropic-compatible APIs so it plugs straight into existing coding tools.
The llama-server dependency is gone
The most deliberate change: 0.2.15 no longer leans on an external llama-server. Under 0.2.14, a model added via "Use a local GGUF" could load through any llama-server binary sitting around your machine, pulled from STRATA_LLAMA_SERVER, your PATH, or Homebrew bins — and the app would quietly relaunch it on next start. Now 0.2.15 loads models only on Strata's own engine, and the model shows an explanatory message instead. That's a deliberate move to remove a moving part, and it removes one that was quietly version-drifting underneath the app.
Drafting gets two fixes at once. Setting it to Off no longer downloads the ~5 GB MTP draft layer, which previously happened anyway,on Mac included, which is a waste you don't need. And a drafting choice made on the Models screen now survives a restart. In 0.2.14 the app silently rewrote it back to Auto, so your Off setting would reload with the DFlash2 drafter regardless. It tested this on a Tesla P100 with an IQ3_XXS quant: Off ready in 104 seconds, DFlash2 in 113, MTP in 116, and the choice held on every restart.
The "no internet, draft layer can't be downloaded" warning now appears only when a model actually starts without a drafter. Under 0.2.14 it could linger long after a drafter was found.
The Qwen3.6-35B-A3B first-run install is finally fixed. In 0.2.14 the download died after 4 seconds. On a clean Linux box with two RTX A6000s, 0.2.15 pulled a 23.75 GB UD-Q4_K_XL model in 320 to 340 seconds to 100%. The progress bar no longer stalls at 99%, the image-reader portion now reflects actual arrival, and the setup screen tells you explicitly when the engine isn't ready yet.
The terminal gets smarter about your hardware
The strata CLI keeps a chosen drafter, MTP, DFlash2, Off, or Auto, across restarts, tested over 8 starts on Strix Halo and a Tesla P100. Every start ran whatever you last picked.
A hybrid NVIDIA+AMD system previously "picked nothing." A machine with both an RTX 3080 and an RX 6700 XT would get "the feed has no engine package for this PC." Now the CLI reads the AMD card, and switching engines reports correctly instead of flagging the other card's engine as not-for-this-computer.
And when an engine won't start because Linux is missing a library, the update now names it and tells you how to install it — sudo apt install libatomic1, for example. That's a real upgrade from the old "the engine program did not run on this computer," which said nothing actionable at all.
Head here to the GitHub release page for more information.
