Strata 0.2.16: Strata's own engine code, a new folder layout, and a long list of fixes
ダウンロード
- install · manifest.txt (0 MiB)
- install.ps1 (0 MiB)
- install.sh (0 MiB)
- latest.json (0 MiB)
- qwen3.8 · flash · next · dflash2.strata1.q4_0 · bf16sel.pack (498 MiB)
- qwen3.8 · flash · next · dflash2.strata1.q4_0 · selq4.pack (324 MiB)
- SHA256SUMS.txt (0 MiB)
- linux · x86_64 (286 MiB)
- linux · x86_64.AppImage.sig (0 MiB)
- macos · arm64 (154 MiB)
- NVIDIA · CUDA 13 (104 MiB)
- NVIDIA · CUDA 13 (0 MiB)
- Control · arm64 (75 MiB)
- Control · x86_64 (90 MiB)
- Control · arm64 (43 MiB)
- Control · x86_64 (64 MiB)
- dsv41 · arm64 · CUDA · CUDA 13 (529 MiB)
- Engine · arm64 · CUDA · CUDA 13 (554 MiB)
- NVIDIA · CUDA 12 (732 MiB)
- Engine · x86_64 · CUDA · CUDA 13 (553 MiB)
- AMD · HIP (258 MiB)
- AMD · HIP (184 MiB)
- AMD · HIP (264 MiB)
- AMD · HIP (609 MiB)
- Engine · x86_64 · sycl (179 MiB)
- Engine · arm64 · Metal (3 MiB)
- NVIDIA · CUDA 12 (705 MiB)
- Engine · x86_64 · CUDA · CUDA 13 (465 MiB)
- AMD · HIP (160 MiB)
- AMD · HIP (210 MiB)
- AMD · HIP (512 MiB)
- feed · aarch64.json (0 MiB)
- feed · aarch64.json.sig (0 MiB)
- feed.json (0 MiB)
- feed.json.sig (0 MiB)
- Strata.app (153 MiB)
- Strata.app.tar.gz.sig (0 MiB)
リリースノート
# Strata 0.2.16: Strata's own engine code, a new folder layout, and a long list of fixes
The app and the Terminal are 0.2.16. The engines are 0.1.62 on every platform, and so is Control, the small program that installs and updates them (there is no engine 0.1.61 release; Control and the engines share one number from here). The update feed is sequence 128.
What changes for you
- Update as usual (see Updating below). Your models, chats and settings stay.
- The engines are new downloads:
strata updateand the app fetch them. Every platform has one (Windows, Linux, Mac, DGX Spark). - Mac M1, M2, M3 or M4 with UD-Q4_K_XL: update. 0.2.15 gave wrong results there (next section).
- Two ways Strata could stop with no message are fixed: a crash at start on some Windows accounts, and Ubuntu's memory guard ending a model load (reduced, not removed: see Known problems).
- On Windows, typing
stratain a default PowerShell works now. - If you build Strata from a clone: the top-level folders moved into
engine/,app/,terminal/,docs/andtools/. Build the engine withcmake -S engine. - New settings are off by default.
Mac M1 to M4: 0.2.15 gave wrong results with UD-Q4_K_XL (fixed, #4141)
On an M1, M2, M3 or M4, 0.2.15 computed only a quarter of the columns in some layers (160 of 640 in the layer we measured). The text still read fluently, so a short answer does not show it. The M5 was not affected. Against the same Mac's default path, on 6 prompts of 512 to 8,192 tokens, the first-token difference (KL) fell from 0.303 to 0.0068, and the top token differed on 1 of 6 prompts instead of 3 of 6. We measured this on an M5 Max forced onto the older kernels, because we had no M1 to M4 Mac. Fix by macintog.
What is new
- Strata's own engine code. The engines for NVIDIA, AMD, Intel Arc and the Mac run on Strata's own code, with no ggml and no llama.cpp matrix kernels. The picture reader is Strata's own too: its file went from 53,864,888 bytes to 807,976 bytes on NVIDIA.
- The answers did not change. On our 10 test prompts (chat, code, a tool call, long pastes, Chinese, German, a picture), the new engines gave the same tokens as 0.2.15's on the RTX PRO 5000 (9 of 9), Tesla P100 (10 of 10), Radeon R9700 (10 of 10), Arc A750 (9 of 9), M5 Max (10 of 10) and Strix Halo (9 of 9 text prompts). The exceptions: the Strix Halo picture answer (intended change in the picture reader), 4 short prompts on the DGX Spark (small number differences, same tokens), and the Windows CUDA 12 engine on an RTX 5070 (2 of 10 answers differ, because it now has native code for that card).
- Faster long prompts on older NVIDIA cards. Tesla V100, a 26,000-token prompt: 17,464 ms with 0.1.58 and 15,068 ms now. RTX 2080 Ti at 26,000 tokens, speed against 0.1.58: 0.885 in 0.2.15 and 1.088 now.
- The chat shows what the model is doing ("Thinking…", "Calling read…", "Writing file app.js…"). A tool call that cannot run no longer ends the turn: the agent gets an error that lists its tools and goes on (20 of 20 in our live check).
Fixed
Windows
- A crash at start on accounts that Windows refuses a system query (WMI), such as a standard user over ssh, is fixed. The background program, the server and the engine ended about 5 seconds after start. On Strix Halo, 3 of 3 starts crashed before; after the fix a full
strata launchserved after 73 s with a chat, a tool call and a picture. A normal desktop user on an RTX 5070 PC was not affected. 0.2.15 has the crash too. Issue #4200 may be the same cause; we could not check it. strataworks in a default PowerShell. Before, PowerShell refusedstrata.ps1("running scripts is disabled"). An older install is fixed on its firststratacommand after the update.install.ps1no longer rewrites your user PATH. It added Strata's folder but changed the type of the value and dropped the last;.UPDATE.batof an old clone ends with exit code 0. Before, it printed "The batch file cannot be found." twice and exited with 1.- After a failed model load, the app's Settings > Updates buttons work again.
Linux
- Ubuntu's memory guard (systemd-oomd) could end
strata launchin a normal terminal while a big model loaded. The window and the engine vanished. It ended 3 of 3 loads in our first round and 2 of 2 in the second, on a Tesla P100 PC with 62 GiB of RAM. Strata now frees the memory that still holds the model's old file copies before the engine starts, and pauses it if memory gets tight. With other programs open, 0 of 5 loads survived without it and 6 of 7 with it (3 of 4 with 12 GiB of other programs open). If it still happens, the nextstrata launchsays what happened. setup.shon a Tesla P100 used to compile the engine (6.5 minutes). It now downloads the ready-made one.- The installer names every card by its own name ("2 x NVIDIA RTX A4000", "AMD Radeon AI PRO R9700"). Before: "NVIDIA NVIDIA RTX A4000", one card named on a two-card PC, or a raw PCI line.
- A false warning on a Tesla P100 ("built for sm_0, sm_1 ... but a graphics card in use is sm_60") is gone.
setup.shwith--gguf-dirno longer asks which model; it picks the same drafter as the Terminal; its tuning question defaults to no and says so;setup.sh --updateprints one line instead of about 15,000 lines of git's list.
Mac
- The app and engine now need macOS 15.0. The public 0.2.15 files needed macOS 27.2.
- The engine no longer stops itself after 60 seconds during long replies in two or more chats at once.
- A UD-IQ1_M model that answered with garbage text is fixed. Metal memory no longer grows between requests (16.9 GB over 4 rounds before, 0.23 GB now).
The Terminal (`strata launch`)
- The Save screen lists what will be downloaded, with sizes, and says "Nothing needs to be downloaded" only when that is true. Offline, it lists what is missing and how to put it in place.
- A model that does not fit the free disk reads "not enough disk space". Before, files in your own folder were marked too (5 of 5 Linux PCs).
- Tuning starts on Skip (before: Quick, about 20 minutes before the first answer). On the Advanced step, Enter goes on and
eedits. - When port 8080 is taken, press
uto use 8081. - The end of a big download says "Finishing the last pieces (310 MB left, 0.4 MB/s)", and the bar no longer goes backwards.
The app
- First run: "Already on this computer" for model files that are there (it said "58.5 GB Download size"). "Checking what your computer can run…" replaces a bare "Connect engine". The first look at the PC took 23 s on an RTX 5070 PC and takes about 2 s.
- The same chat open in two windows now shows the running reply in both.
- The app's local
/v1address accepts any model name when one model is loaded (it answered HTTP 400). - Settings > Updates no longer ends with "The app updater is not ready yet" after a finished update.
- On NVIDIA Blackwell cards, the CUDA 12 engine's first answer took 173.0 s and now takes 0.34 s (#5784).
Known problems
- Windows, Strix Halo and PCs with a small page file: the engine can stop with no message (exit code 0xC0000409) when Windows' memory limit (memory plus page file) runs out. The model's card now shows a notice after the model has loaded. What to do: make the Windows page file larger, choose a shorter context, or use
--expert-cache 8192in the Terminal (it left 18.4 GB free in our test). The same in 0.2.15 (#1607). - Windows from 0.2.15 after the crash above: one `strata update` can leave the engine behind. It says "Engine 0.1.60 NOT changed ... Run
strata update --force". What to do: runstrata updateagain. In the app: press the Strata row in Settings > Updates, then the Engine row. - Ubuntu and other desktops with the memory guard: a big model load can still be ended when little memory is free (3 of 4 loads survived with 12 GiB of other programs open). What to do: close other programs (a web browser uses a lot), or start Strata f