What changed in Strata after October 12? Reading the updates from October 12 to 17 (v0.1.43 to 0.1.44.1)

2026年10月18日

Strata is an inference engine (the base program that runs an AI model) that has become a talking point for running huge models on a 12GB graphics card. After I wrote my article on how it works on October 12, Strata was updated further. Which parts of that article no longer hold?

In this article I reread the official release notes and documents and sorted out what has changed since then. My article where I ran Strata and measured it is here. I did not run any new measurements this time. This reflects what I checked around 8 a.m. on October 17, 2026.

Method: scope and sources

I looked at the release history of Niko1221/Strata on GitHub (v0.1.43.1, v0.1.44 and v0.1.44.1), the README, and the docs MODELS, MAC and STRIX_HALO. I counted the number of stars and pull requests with the GitHub API. All the speed and quality numbers are the author’s published values, and I have not verified them on this blog.

Result 1: how the numbers moved

Result 2: where my previous article no longer matches

Three points I wrote in the previous article differ from the current official explanation.

For Strix Halo chips such as the Ryzen AI Max+ 395 (the same family of chip as the EVO-X2 used on this blog), I wrote “experimental" last time. It is still “experimental," but a dedicated document (STRIX_HALO) has been added, and it says the author measured on the maintainer’s own machine with a Ryzen AI Max+ 395 and 128GB of memory.

Result 3: quality by quantization (the author’s measurements)

“Quantization" is a method of making a model smaller by making its numbers coarser. The smaller it gets, the faster and lighter it becomes, but the more its answers tend to drift from the original. Using Q8_0 (8-bit) as the reference, the author measured how often “the next word chosen is the same as the reference" at 4,995 positions in text.

The trade-off with speed is also published. On the author’s AMD R9700 (32GB), with UD-Q4_K_XL and all experts kept in memory, writing a story runs at 63 tok/s and code at 85 to 88 tok/s. With Q8_0 they were 53 to 57 tok/s and 66 to 74 tok/s. The author writes that llama.cpp on the same machine runs at 16 to 20 tok/s.

Result 4: a change in 0.1.43 backfired and was reverted

In v0.1.43, “adaptive overlap" (overlapping processing steps) was turned on by default for NVIDIA on Linux. However, on some models the writing speed dropped. In the author’s re-measurements, an RTX 5070 Ti desktop ran at 68.4 tok/s (86.4 tok/s with it off), the RTX 5060 Ti was 19% slower, and the RTX 5090 Laptop was 21% slower. So v0.1.43.1 turned the default back off. The RTX 5060 Ti is the model I measured in the Qwen3.8-Flash-Next article. I have not checked how this change relates to the Strata version I used when I measured 70.7 tok/s there.

Result 5: bugs fixed in v0.1.44.1

• A problem where startup stopped without showing anything on old CPUs without AVX2 (such as the Xeon X5670)
• A problem where the engine could not be built in CUDA 12.0 / 12.1 environments
• A problem where answers broke on the Apple M2 (the Metal code exceeded the limit on the number of usable threads)

Discussion: how my previous article reads now

The conclusion of the previous article was “Strata is an engine dedicated to Qwen3.8-Flash-Next." Since it now also supports Qwen3.6-35B-A3B, calling it “dedicated" is too strong. It is still limited to Qwen-family MoE models (models that switch between specialist experts), and that has not changed.

The quality table shows that the coarser the quantization, the lower the agreement with the reference. In the author’s measurements, going from UD-IQ4_XS (88.9%) to UD-Q4_K_XL (93.1%) raises agreement by 4.2 points. However, the memory required also grows (about 88GB for UD-Q4_K_XL). One way to see it is that the options for machines with 128GB of memory have increased.

Limitations: what this article cannot say

• All the speed and quality numbers in this article are the author’s published values. They are not measurements made on this blog.
• The Strix Halo and Mac numbers come from one or two of the author’s machines.
• Strata is updated many times a day. This article, too, reflects the time I checked.

Summary: three points fixed, and more ways to choose quantization

Between October 12 and 17, Strata gained support for Mac and for Qwen3.6-35B-A3B, and you can now choose Unsloth’s 4-bit to 8-bit quantizations. The change in 0.1.43 made some NVIDIA models slower, so it was reverted in 0.1.43.1.

If you already use Strata, update with “UPDATE.bat" (./update.sh on Linux), and measure the speed again on your own machine after updating. If you are trying it for the first time, read “What to know before using it" in the previous article, and the official INSTALL and SECURITY documents, first.

Sources

• Niko1221/Strata (GitHub) release history v0.1.43.1, v0.1.44 and v0.1.44.1
• Strata docs: MODELS, MAC and STRIX_HALO
• This blog: Strata, how it works (article 18195)
• This blog: the article where I ran Qwen3.8-Flash-Next (article 17994)