Strata replaced its GPU and Mac model engines with in-house code and fixed incorrect calculations on some Apple silicon chips. Developers using local models can update to get the new engines while keeping models, chats, and settings.
1. A new engine without external compute kernels
The app and terminal are version 0.2.16, while platform engines and the installer/updater are version 0.1.62. The new engine uses in-house code instead of external matrix computation code.
Image reading also switched to an in-house implementation. According to Strata, the NVIDIA image reader file shrank from 53,864,888 bytes to 807,976 bytes.
Strata compared output tokens with the previous engine using 10 prompts covering chat, code, tool calls, long inputs, Chinese, German, and images. In its tests, Tesla P100 and Radeon R9700 each matched tokens on all 10.
Strata said the Mac M5 Max also matched tokens on all 10 prompts. This shows whether answers were preserved in those tests after changing the execution code.
In Strata's tests, RTX PRO 5000 and Arc A750 each matched tokens on all 9 prompts compared. However, Strata said the Windows CUDA 12 engine gave different answers on 2 of 10 prompts on RTX 5070, which received dedicated code.
According to Strata, Tesla V100 processing time for a 26,000-token prompt fell from 17,464 milliseconds to 15,068 milliseconds. This was about 13.7% less time than engine 0.1.58 for an input of the same length.
The same announcement put RTX 2080 Ti's relative speed for 26,000 tokens at 1.088, up from 0.885 in the prior release, against engine 0.1.58. This is about 22.9% faster than release 0.2.15, a different metric from V100's time reduction.
Table: Long-input processing performance before and after
Conditions: Strata's reported results, 26,000-token input; V100 compared with engine 0.1.58; RTX 2080 Ti compared across releases using speed relative to 0.1.58
Units: V100 in milliseconds; RTX 2080 Ti in relative speed (dimensionless)
| Hardware | Previous comparison baseline | Before | Current |
|---|---|---|---|
| Tesla V100 | Engine 0.1.58 | 17,464 | 15,068 |
| RTX 2080 Ti | Release 0.2.15 | 0.885 | 1.088 |
Chart: Tesla V100 processing time. Conditions: Strata's reported results, 26,000-token input; lower means shorter processing time. Unit: milliseconds
xychart-beta
x-axis ["Engine 0.1.58", "Current"]
y-axis "Processing time (milliseconds)" 0 --> 18000
bar [17464, 15068]Chart: RTX 2080 Ti relative speed. Conditions: Strata's reported results, 26,000-token input, speed relative to engine 0.1.58; higher is faster. Unit: relative speed (dimensionless)
xychart-beta
x-axis ["Release 0.2.15", "Current"]
y-axis "Relative speed" 0 --> 1.2
bar [0.885, 1.088]2. Fixing layers that calculated incorrectly on M1–M4
The prior version computed only a quarter of the columns in some layers of UD-Q4_K_XL models, which store weights at low precision. One measured layer computed 160 of 640 columns. Natural-sounding short replies hid the error.
Strata said the fix targets M1–M4 and that M5 was unaffected.
Strata measured the fix by forcing older compute kernels (low-level computation code) on an M5 Max. KL measures the difference between two predictive probability distributions. Here it compares first-token probabilities with the same device's default path.
Table: Effects of fixing the older Mac computation path
Conditions: Strata's reported results, older kernels forced on M5 Max, 6 prompts of 512–8,192 tokens, compared with the same device's default path
Units: KL difference is dimensionless; top-token mismatches are counted in prompts
| Metric | Before fix | After fix |
|---|---|---|
| First-token KL difference | 0.303 | 0.0068 |
| Top-token mismatches | 3 | 1 |
Under the same conditions, both the probability difference and the number of prompts with a different highest-probability token fell. This case showed the importance of directly comparing calculations rather than natural-sounding answers.
The minimum operating system for the Mac app and engine changed to macOS 15.0. Previously published files required macOS 27.2.
Meaningless character output from UD-IQ1_M models was also fixed.
According to Strata, Metal (Apple's graphics and compute technology) memory growth between requests over 4 rounds fell from 16.9GB to 0.23GB.
3. Mitigating startup failures and memory-related shutdowns
Windows no longer exits about 5 seconds after launch on some accounts lacking system query permissions. Strata said Strix Halo failed all three startup tests before the fix; afterward, a chat, tool-call, and image run responded in 73 seconds.
The issue preventing the strata command from running in the default PowerShell was also fixed.
To address Ubuntu's memory protection terminating runs while loading large models, the update first frees memory occupied by previous copies of model files and pauses the engine when memory is insufficient.
According to Strata, a Tesla P100 PC with 62GiB of memory failed all 5 attempts with other programs open before the fix, but succeeded in 6 of 7 afterward. With other programs using 12GiB, it succeeded in 3 of 4 attempts.
Preparation time also fell. According to Strata, the app's initial system check on an RTX 5070 PC dropped from 23 seconds to about 2 seconds.
Strata said the CUDA 12 engine's first response on NVIDIA Blackwell fell from 173.0 seconds to 0.34 seconds.
4. Check memory headroom and hardware test coverage
On Windows, Strix Halo or PCs with small page files (disk space used as memory) may silently exit at the combined memory limit. Ubuntu systems may also stop loading models when memory is low. Actual older Apple silicon devices and some platform changes remain untested.
5. Get the new engine while keeping existing data
Runtime stability fixes are separate from the switch to in-house code. Updating requires downloading the new platform engine. Tesla P100 installation on Linux now downloads a prebuilt engine instead of compiling locally, reducing build work.
Regular terminal users update with the following command. Windows users who encountered the previous version's startup crash and whose engine has not updated are advised to run the same command once more.
strata updateApp users can use the built-in updater or a new installer. Developers building from a repository clone must account for the engine path moving to engine/.
Strata 0.2.16 strengthens the foundations of local execution by switching to an in-house engine and fixing calculation errors.
