Model card for Sync-Strata 0.1.1 (quality, speed, support section)
Browse files
README.md
ADDED
|
@@ -0,0 +1,113 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
license_name: qwen-community-1.0
|
| 4 |
+
license_link: LICENSE
|
| 5 |
+
base_model:
|
| 6 |
+
- Qwen/Qwen3.8-Flash-Next
|
| 7 |
+
base_model_relation: quantized
|
| 8 |
+
tags:
|
| 9 |
+
- qwen3.8-flash-next
|
| 10 |
+
- moe
|
| 11 |
+
- sync-strata
|
| 12 |
+
- experimental
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# Qwen3.8-Flash-Next Sync-S (experimental)
|
| 16 |
+
|
| 17 |
+
A re-quantized expert pack of **Qwen3.8-Flash-Next** for one consumer NVIDIA GPU, made to run faster than upstream
|
| 18 |
+
Strata's Gyro. It runs only on **[Sync-Strata](https://github.com/hiruhiru2005/sync-strata)** (a fork of
|
| 19 |
+
[Strata](https://github.com/Niko1221/Strata)). **This is not a GGUF: it does not load in llama.cpp or upstream Strata.**
|
| 20 |
+
|
| 21 |
+
Recommended for: **PCs with 32 GB of RAM**.
|
| 22 |
+
|
| 23 |
+
## What was done
|
| 24 |
+
|
| 25 |
+
- The routed experts were re-quantized to a trellis format of the Gyro / QTIP family, with the bit width chosen per
|
| 26 |
+
expert. The pack also carries the dense weights that the engine reads from it.
|
| 27 |
+
- The remaining weights (embeddings, the n-gram/PLE table and so on) are not in this repository: they are in the
|
| 28 |
+
shared parts, [hiru2005/Sync-shared](https://huggingface.co/hiru2005/Sync-shared), which every Sync model uses.
|
| 29 |
+
|
| 30 |
+
## Files
|
| 31 |
+
|
| 32 |
+
Total: **28.14 GB (26.21 GiB)**, plus the shared parts.
|
| 33 |
+
|
| 34 |
+
| File | Size | What |
|
| 35 |
+
|---|---:|---|
|
| 36 |
+
| `MANIFEST.json` | 1,835 B | Index of the pack (file names, sizes, checksums) |
|
| 37 |
+
| `compat-bf16.json` | 36,116 B | Layout and conversion tables read by the engine |
|
| 38 |
+
| `conversions.json` | 151,849 B | Layout and conversion tables read by the engine |
|
| 39 |
+
| `dense_nq.txt` | 2,013 B | Layout and conversion tables read by the engine |
|
| 40 |
+
| `experts_k.txt` | 358,331 B | Layout and conversion tables read by the engine |
|
| 41 |
+
| `hadamard.txt` | 8,938 B | Layout and conversion tables read by the engine |
|
| 42 |
+
| `index.txt` | 93,669 B | Layout and conversion tables read by the engine |
|
| 43 |
+
| `native_experts.txt` | 3,543 B | Layout and conversion tables read by the engine |
|
| 44 |
+
| `r1q2.txt` | 3,818 B | Layout and conversion tables read by the engine |
|
| 45 |
+
| `dense_nq.bin` | 755,548,160 B | Dense weights read by the engine from the pack (second part) |
|
| 46 |
+
| `dense.bin` | 1,538,035,200 B | Dense weights read by the engine from the pack |
|
| 47 |
+
| `experts.v2.bin` | 25,846,112,256 B | Routed experts (re-quantized) |
|
| 48 |
+
| `LICENSE` | | Qwen Community License 1.0 |
|
| 49 |
+
|
| 50 |
+
## Requirements
|
| 51 |
+
|
| 52 |
+
- Windows 10/11 or 64-bit Linux. Supported GPUs: NVIDIA RTX 30/40 series (driver R580 or newer); RTX 20, GTX 16, RTX 50 are expected to work, and GTX 10, Radeon RX 6000/7000/9000 and the integrated Radeon of Ryzen AI Max (Strix Halo: 8060S, 8050S, 8040S; Windows and Linux) are experimental, all not checked on real hardware. Other integrated GPUs and Intel GPUs are not supported ([the GPU table](https://github.com/hiruhiru2005/sync-strata#supported-gpus)). A CPU with AVX2.
|
| 53 |
+
- RAM and VRAM: see "Recommended for" above. The app picks the model by the RAM installed in the PC; if the free RAM is short it warns you instead of choosing a smaller model. The minimum amounts of RAM and VRAM are to be measured.
|
| 54 |
+
- Disk: this pack plus the shared parts (`native-gguf`, `mtp-rt`, `tokenizer`, `mmproj` (optional)). An NVMe SSD is recommended.
|
| 55 |
+
|
| 56 |
+
## Use with Sync-Strata
|
| 57 |
+
|
| 58 |
+
The Sync-Strata app downloads this model for you: open the app and pick **Sync-S** on the first screen (or let
|
| 59 |
+
"auto" choose). To download by hand, put the files in a folder named `sync-s-sb` inside the app's models folder (`models_dir` in `settings.json`), together with the shared parts.
|
| 60 |
+
|
| 61 |
+
```
|
| 62 |
+
sync settings set model sync-s
|
| 63 |
+
sync up
|
| 64 |
+
```
|
| 65 |
+
|
| 66 |
+
## Quality
|
| 67 |
+
|
| 68 |
+
Measured on Colab with the production eval_v4 (8 domains, 1,400 windows; the teacher is the Q8_0 GGUF of the original model).
|
| 69 |
+
|
| 70 |
+
| | Sync-S |
|
| 71 |
+
|---|---:|
|
| 72 |
+
| KL to the original model (all positions, lower is better) | 0.2310 |
|
| 73 |
+
|
| 74 |
+
This value was measured on the earlier form of Sync-S (table-lookup points, the same dense part). The form in this repository (no table lookup) has not been measured yet; for Sync-XS and Sync-M the two forms were on par, within the measurement noise.
|
| 75 |
+
|
| 76 |
+
For comparisons with other models (upstream Gyro-S, Gyro-M, IQ3_XXS, Sync-XS), see the [Sync-XS page](https://huggingface.co/hiru2005/Sync-XS).
|
| 77 |
+
|
| 78 |
+
## Speed
|
| 79 |
+
|
| 80 |
+
Measured on the author's PC with Sync-Strata 0.1.0 (Ryzen 7 7800X3D, 32 GB DDR5, RTX 3080 12 GB, NVMe SSD; Windows 11 and Linux). The values in the table are for Windows; the Linux values are to be measured. No numbers for GTX 10 or Radeon: no hardware to measure on.
|
| 81 |
+
|
| 82 |
+
| | tail skip | jbench decode (tok/s) | dsh chat (tok/s) |
|
| 83 |
+
|---|---|---:|---:|
|
| 84 |
+
| Sync-S | R4 | 144.0 | 104.1 |
|
| 85 |
+
| Sync-S | R7 | 127.5 | 90.7 |
|
| 86 |
+
| Sync-S | R10 | 104.6 | 80.1 |
|
| 87 |
+
|
| 88 |
+
Speed depends strongly on the PC. The values of the other models, including upstream Gyro, are in the table of the [Sync-Strata README](https://github.com/hiruhiru2005/sync-strata#benchmarks).
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
## Calibration
|
| 92 |
+
|
| 93 |
+
The quantization was calibrated on public text (code, documentation, encyclopedia articles, several languages, agent
|
| 94 |
+
traces) and on some text made from the author's own chat logs. The calibration text is not distributed.
|
| 95 |
+
|
| 96 |
+
## Credits
|
| 97 |
+
|
| 98 |
+
- Base model: [Qwen/Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) by the Qwen team.
|
| 99 |
+
- [agentionai](https://huggingface.co/agentionai): the Gyro trellis format and kernels, and the
|
| 100 |
+
[Gyro-S GGUF](https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF) that the shared parts are cut from.
|
| 101 |
+
- [Strata](https://github.com/Niko1221/Strata) by Niko1221 and the Strata contributors (MIT), which Sync-Strata forks.
|
| 102 |
+
- QTIP (Tseng et al., 2024); [llama.cpp / ggml](https://github.com/ggml-org/llama.cpp) (MIT).
|
| 103 |
+
|
| 104 |
+
## Support (optional)
|
| 105 |
+
|
| 106 |
+
Sync-Strata and the Sync models are free. A donation page is at [Ko-fi](https://ko-fi.com/hiruhiru2005): a thank-you only, with nothing in return (no perks, no priority support, no feature promises), and it does not change the license terms.
|
| 107 |
+
|
| 108 |
+
## License
|
| 109 |
+
|
| 110 |
+
These files are derived from Qwen3.8-Flash-Next and are under the **Qwen Community License 1.0** (`LICENSE`,
|
| 111 |
+
Copyright (c) 2026 Qwen). Keep `LICENSE` with every copy. Commercial products with more than 100 million monthly active
|
| 112 |
+
users or US$20 million monthly revenue must show the model name, and a Model-as-a-Service or AI work assistant business
|
| 113 |
+
needs a separate license from Qwen for commercial use. Read the full text.
|