Sync-Strata Fits 125B in 32 GB. Its Top Speed Skips Experts

A 125B model on a 32 GB PC, and the setting that makes it fast

Sync-Strata runs Qwen3.8-Flash-Next, a 125B MoE model, on a 32 GB Windows PC at 144.0 tokens a second. That is the headline on the project's page, and it is measured on the author's own machine: a Ryzen 7 7800X3D, 32 GB of DDR5, an RTX 3080 12 GB and an NVMe drive.

The number is real for the setting it names. It is not the whole story. The engine's speed setting is called tail skip, and R4, the setting behind the 144.0, uses fewer experts per token to get there. R10 skips none and is the quality-first setting. On the same PC, Sync-S at R10 writes 104.6 tokens a second, not 144.0.

That gap is the story: a fork of Strata with its own quants, built for PCs that upstream Strata's Gyro models do not fit well, and a speed dial that trades quality for throughput. The project says so in its own settings table. The article should say it too.

What Sync-Strata is

Sync-Strata is a fork of Strata by @hiru_2005, at version 0.1.0; the project does not name the upstream version it forked. It runs the same Qwen3.8-Flash-Next model as Strata, but with three re-quantized expert packs of its own: Sync-XS, Sync-S and Sync-M. The packs are in a trellis format of the Gyro and QTIP family, with a bit width chosen per expert. They are not GGUF, and they do not load in llama.cpp or in upstream Strata.

The app is a desktop program for Windows 10 and 11 and 64-bit Linux. It ships its own chat client, dsh, and serves an OpenAI-compatible API on the local machine. You can use it from the desktop, a phone on the same Wi-Fi, the command line or another program. It uses one GPU, the supported card with the most memory, and it needs an AVX2 CPU, an NVIDIA driver from R580 and at least 8 GB of VRAM.

The project labels itself experimental, and the format warning matters: a Sync pack is tied to this engine. There is no dropping it into your existing Strata or llama.cpp setup.

Which model for which RAM

RAM in the PC Model Pack size KL to the original model
24 GB or less Sync-XS 24 GB class 0.2851
32 GB Sync-S 28.14 GB 0.2310
48 GB or more Sync-M 34.59 GB 0.1642

Each pack also needs shared parts, which the README puts at about 35 GB of disk. The app and the packs are separate downloads, and an NVMe drive is recommended.

The KL column is the distance from the original model, lower is better. Two caveats sit inside it. Sync-S's 0.2310 was measured on the earlier form of the pack; the form in the repository has not been measured, and the author expects it to be on par. Sync-M's 0.1642 is the current form. Sync-XS's 0.2851 is the current form and the roughest of the three, which is what a 24 GB PC buys.

The speed table

All of these are the author's numbers on that one PC, Windows 11, engine 0.1.0. jbench is upstream Strata's benchmark: 10 fixed requests, decode as a geometric mean, the median of three runs. dsh chat is six casual Japanese chats, two to three runs.

Model Tail skip jbench decode jbench prefill dsh chat
Sync-S R4 144.0 597 104.1
Sync-S R7 127.5 648 90.7
Sync-S R10 104.6 593 80.1
Sync-M R4 124.6 316 92.2
Sync-M R7 93.6 322 75.7
Sync-M R10 69.1 322 59.1
Gyro-S, upstream, default 7 72.6 241 not measured
Gyro-M, upstream, default 7 43.2 93 not measured
Gyro-M, upstream, fastest 4 67.8 102 not measured

The comparison the project wants you to see is Sync-S R4 at 144.0 against Gyro-S at 72.6, or Sync-M R4 at 124.6 against Gyro-M at 43.2. Those are large gains. They are also comparisons of different quantizations on a different engine, and the Sync side is at its fastest, most aggressive setting.

The quality-first comparison is smaller. Sync-S R10 at 104.6 against Gyro-S at 72.6 is about 44% faster, not double. Sync-M R10 at 69.1 against Gyro-M's fastest 67.8 is level. That is the pair to use when quality is the constraint.

The tail skip, said plainly

The setting has four choices: Auto, R4, R7 and R10. R4 uses the fewest experts per token and is the fastest. R10 skips none and is the highest quality. Auto picks R4 or R7 for your PC. The README says this in one line, and the model cards repeat the same table.

What the project does not publish is a quality number for each setting. The KL figures are one per pack, not one per R4, R7 and R10. So the cost of R4 is not a measured KL in the release. What we can say is structural: R4 computes fewer experts, so it is a different, cheaper model at run time, and the 144.0 headline belongs to that setting. R10 is the one that computes every expert the pack expects.

That is why the speed table matters more than the headline. A reader who wants the quality the KL number describes should look at R10. A reader who wants the fastest chat should know they are choosing R4 and that the project has not measured what it costs in KL.

The quality ladder

The model cards compare Sync-XS with the alternatives on the same evaluation, scale B:

Sync-XS against Result
Gyro-S, upstream 16.1% better
IQ3_XXS 48.7% worse
Gyro-M, upstream 39.3% worse
Sync-S 34.8% worse
Sync-M 102.2% worse

So the 24 GB model is better than the old Gyro-S and worse than everything above it. Sync-S sits between XS and M, and Sync-M is the closest to the original. That is the honest ladder: the less RAM you have, the lower the quality, and the numbers say how much lower.

The pack-level KL values are 0.2851 for XS, 0.2310 for S and 0.1642 for M. A lower KL is better. The jump from XS to S is the one a 32 GB buyer is paying for; the jump from S to M is the one a 48 GB buyer is paying for.

What the comparison does not settle

One PC, one GPU, one operating system. The Linux values are not measured. The GPU table checks RTX 30 and 40 cards; RTX 20, GTX 16 and RTX 50 are expected to work but not checked, GTX 10, V100, Titan V and Radeon are experimental, and Intel and other integrated GPUs are not supported. Macs are not supported.

The speed comparison is not a controlled engine test. The Sync models are different quants from Gyro, the Gyro runs used upstream Strata 0.2.13.1 while the Sync side is a fork with its own changes, and the author ran both sides. The jbench medians of three runs and the dsh chat runs are more than a single sample, but they are still one machine and one author's harness.

The quality numbers are Colab evaluations against the Q8_0 GGUF of the original model, with 1,400 windows across eight domains. The tail-skip settings are not separated in them. The Sync-S number is from the earlier pack form. Those are real limits, and the project states most of them.

Who it is for

It is for a Windows or Linux PC with 32 or 48 GB of RAM and an RTX 30 or 40 card, where upstream Strata's Gyro models are too large or too slow, and where a desktop app with a phone client and an OpenAI-compatible API is useful. The 32 GB / Sync-S combination is the one the project is built around: 28.14 GB of experts plus the shared parts, in a machine that would otherwise be choosing between a much smaller model and a much slower one.

It is not for a Mac, an Intel GPU, a 24 GB PC that wants the best quality, or anyone who needs the pack to load in llama.cpp or upstream Strata. The format is deliberately its own.

The verdict

The 32 GB angle is real, and it does not depend on skipping: Sync-S runs on the 32 GB PC at R10, every expert computed, at 104.6 tokens a second. The 144.0 is the same PC at R4, and the quality ladder gives a 24 GB and a 48 GB buyer a clear choice.

The headline needs its setting attached. 144.0 is Sync-S at R4, the setting that skips the most experts. The quality-first Sync-S is 104.6 at R10, and Sync-M at R10 is 69.1. Those are still good numbers for this class of PC, and they are the ones to quote when the reader cares about the model's quality rather than its fastest possible chat.

The project is experimental, the packs are engine-specific, and the quality cost of R4 is not measured. Say all three before quoting the speed.

Sources

FAQ

What is Sync-Strata?

A fork of Strata by @hiru_2005 that runs Qwen3.8-Flash-Next, a 125B MoE model, with its own Sync-XS, Sync-S and Sync-M expert packs. It is a Windows and Linux desktop app with a chat client and an OpenAI-compatible API, and it is experimental.

How much RAM does Sync-Strata need?

24 GB or less for Sync-XS, 32 GB for Sync-S, 48 GB or more for Sync-M. Each pack also needs about 35 GB of shared parts on disk, and an NVMe drive is recommended. The app picks the model by installed RAM; if free RAM is short it warns rather than switching down.

Why is Sync-S 144.0 tok/s in one table and 104.6 in another?

The 144.0 is tail skip R4, the fastest setting, which uses fewer experts per token. The 104.6 is R10, which skips none. The project does not publish a separate quality number for each setting, so the cost of R4 is not measured in KL.

Is Sync-Strata better than Strata?

It is faster on the author's PC at the aggressive settings, with its own quants. The comparison is not like for like: different packs, a forked engine, one machine and the author's harness. Upstream Strata's Gyro-S at its default measured 72.6 tok/s on the same PC, against Sync-S's 144.0 at R4 and 104.6 at R10.

Does Sync-Strata work on a Mac or an AMD card?

Macs are not supported. Radeon RX 6000, 7000 and 9000 cards and the integrated Radeon of Ryzen AI Max are experimental and not checked on real hardware. RTX 30 and 40 are the checked NVIDIA cards.

Can I load a Sync pack in llama.cpp or upstream Strata?

No. The packs are not GGUF and the project says the model and engine formats are not compatible with upstream Strata or llama.cpp.