Strata 0.2.11: New Compressed Model Files, Faster Prompting, and a Privacy Bug Fixed
Strata 0.2.11 shipped today, 5 November 2026, and it's the most feature-heavy release of the project so far. The update brings a new family of ultra-compressed model files, faster prompt processing on both NVIDIA and AMD cards, the ability to start big models across several cheap graphics cards, chats stored as files, and, arguably the most important, a fix for a privacy bug that could send one user's replies to another. It's a lot to pack into what is a two-month-old project.
The engine, built by single developer Niko Veit with a volunteer contributor base, has climbed to roughly 40,000 GitHub stars in under two months. Strata lets you run Alibaba's 125-billion-parameter Qwen3.8-Flash-Next on hardware with as little as 8 GB of VRAM, keeping everything offline. This release sharpens that basic promise in several directions.
New Unsloth files
The headline feature is a set of seven new quantizations of Qwen3.8-Flash-Next from Unsloth, downloadable through the app, the Terminal, or a local folder. Each is verified by size and SHA-256 against Unsloth's revision.
The lightest is UD-IQ1_M at 68.0 GB. The ~4-bit UD-IQ4_XS-v31 tops out near 93.7 GB. That sounds ridiculous against a 12 GB card, so here's why it works: the model uses a Mixture of Experts architecture. It has 125 billion parameters but activates only around 6 billion per token, leaving most of it dormant. You don't need the whole thing in VRAM at once, just the experts being asked for right now, with the rest fetched on demand.
Real decode speeds, from the maintainer's own tables: the 1-bit UD-IQ1_M hits 43–45 tok/s on a DGX Spark, 101–114 tok/s across an RTX A6000, and 31–37 tok/s on a 12 GB RTX 3060. Not desktop-smooth, but genuinely usable.
One thing to watch: the UD-IQ1_S build does not load. A layer there mixes two expert formats Strata can't read, so UD-IQ1_M is your 1-bit choice. The notes also credit a new expert format (Q5_0) that cuts error from roughly 2.6–3.0% down to about 1.2–1.3%.
Faster prompting
This release spends real energy on prompt processing, reading your input, which matters enormously for agent workloads that shove large context at the model.
On NVIDIA cards with compute capability 8.0 or higher, Strata now uses fast compact-integer MMQ kernels instead of converting weights to 16-bit. On an RTX PRO 6000, prompt speed jumped +26% to +38% across 4K–32K prompts. A new --prefill auto mode reads chunks up to 32,768 tokens (previously 8,192), giving up to +70% on a streamed setup.
AMD isn't an afterthought. Four Radeon AI PRO R9700 cards running the all-cards split engine saw combined throughput for eight simultaneous users rise 32.7%, and the change drops a dependency (librccl) that had forced prompts through a single card's memory. A new FP16 weight cache on NVIDIA Linux keeps about 3 GB of card free and adds roughly 4–7% throughput under load.
Multi-card starts that actually work
Strata 0.2.11 fixes something that used to fail hard: starting a model across several small-memory cards. Two 8 GB cards used to throw "too little free VRAM" and stop cold.
The new version walks down a memory list, trimming weights, shrinking the context window in steps, and dropping under-utilized cards out until something launches. The notes claim every start that failed before now answers a chat on a tight budget pair of 8 GB RTX PRO 6000-class cards.
Chats are files now
A structural change worth flagging: each conversation is now saved as a file (chats<id>.json) in Strata's data folder, with byte-for-byte fidelity and no 5 MB limit. Writes go to a temporary file first and are renamed into place, so a crash leaves the last good copy.
If a chat can't be written, it stays whole in memory with a "Copy unsaved chats" button. The privacy line on the settings screen now reads plainly: conversations stay on this computer, in Strata's data folder.
The privacy fix
Of all the changes, the one that draws the most attention is the fix for issue #2892. When a model ran in multiple copies, a reply could go to the wrong request, returning an empty answer, text containing someone else's earlier prompt, or another user's token counts.
The maintainer reproduced 30 of 30 crossed hand-offs on the old code. After the fix, 30 of 30 forced requests and 60 of 60 unforced ones came back clean. It's a Python-server-side bug, which matters if you run the engine as a shared service.
The smaller changes
The rest of 0.2.11 rounds things out. A running agent turn now survives a page reload, showing partial progress with a Stop button instead of "Network error." A Markdown rendering bug that could chew through memory — a suspected cause of a reported 4.4 GB memory blow-up on Windows, was uncovered and fixed while smoothing out streaming previews. Menus now reposition to stay readable at every interface scale, and a malformed Host header that used to crash the host now returns a clean 400.
Security got a light pass too: Pillow moved to 13.0.0 to patch a high-severity JPEG 2000 vulnerability, API-key comparisons now use constant-time checks, image-tool prompts no longer travel on the command line, and only a fixed safe list of image formats is opened.
What to keep in mind
The feature set is impressive, but the versioning is a bit unusual. The app and Terminal are 0.2.11, while Control and every engine package sit at 0.1.57. Keep that in mind if something behaves differently than the release notes suggest.
The changes also come with trade-offs the project documents honestly. NVIDIA K-quant prompt kernels are on by default now (turn off with STRATA_MMQ_KQ=0), the auto-prefill ceiling sits at 32,768 on capable cards, and chats now live in the data folder rather than browser storage.
Performance-wise, the independent evidence backs the core claim: Strata's advantage is largest on small GPUs, which is exactly where it was built for. One tester held to 12 GB found it decoding 2.7× faster than llama.cpp. Quality holds up better than low-bit numbers imply,same-file tests showed essentially no gap versus llama.cpp.
But the limits are real. Quality degrades at 2–3 bits, some agent tool calls still fail, long prompts stay slow on 12 GB cards, and memory pressure can be brutal. The model weights carry the Qwen Community License 1.0, not MIT, so check before commercial use. And Strata ships with an install script you run with user rights, updates roughly every two days, and has no external security audit.
The new features move the floor for local inference, no question about it. But read the self-reported peak numbers, the famous "6× faster than llama.cpp" is mostly prompt speed, with real decode gains closer to 2–3×.
To try 0.2.11, head here to the GitHub release page. The notes include full benchmark tables and the exact environment variables to revert any change, with the technical background in the project's "How Does Strata Work?" doc and its published paper.
