Sync-shared / README.md
hiru2005's picture
Model card for Sync-Strata 0.1.1 (quality, speed, support section)
60e7b9e verified
|
Raw History Blame Contribute Delete
3.64 kB
metadata
license: other
license_name: qwen-community-1.0
license_link: LICENSE
base_model:
  - Qwen/Qwen3.8-Flash-Next
base_model_relation: quantized
tags:
  - qwen3.8-flash-next
  - moe
  - sync-strata
  - experimental

Sync shared parts for Qwen3.8-Flash-Next (experimental)

The parts every Sync model (Sync-XS, Sync-S, Sync-M) needs besides its own pack. Download them once. They are used only by Sync-Strata; the Sync-Strata app downloads them for you.

Files

native-gguf/

Shared weights cut from agentionai's Gyro-S GGUF (embeddings, n-gram/PLE table and the other non-expert tensors the engine reads with --native / --ple-gguf).

The file list and sizes of this folder will be added when it is published. The whole Gyro-S file is 58.49 GB; this folder keeps only the parts the engine reads.

mtp-rt/

The multi-token-prediction (draft) layer, converted from the original BF16 checkpoint for the engine; draft_vocab.bin comes from upstream Strata (MIT).

Total: 0.82 GB (0.77 GiB)

File Size
dense.bin 116,099,072 B
dense.txt 1,880 B
draft_vocab.bin 425,196 B
experts.bin 707,788,800 B

tokenizer/

The tokenizer and chat template of Qwen3.8-Flash-Next.

Total: 0.01 GB (0.01 GiB)

File Size
chat_template.jinja 9,993 B
merges.txt 3,600,844 B
token_type.json 744,960 B
tokenizer.json 554 B
vocab.json 5,737,005 B

mmproj/ (optional)

Optional vision encoder and projector (BF16). Not hosted here: the Sync-Strata app downloads it from its original repository, ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF. Its card says the files inherit the license of the base model (Qwen Community License 1.0); the repository metadata says Apache-2.0.

File Size
mmproj-Qwen3.8-Flash-Next-BF16.gguf (original repository) 907,543,008 B

Where they come from

  • native-gguf is cut from agentionai/Qwen3.8-Flash-Next-Gyro-GGUF (Gyro-S; the revision will be added). It keeps only the tensors the engine reads; the tensors are meant to be byte-for-byte those of the original file, but this has not been checked with a tool yet.
  • mtp-rt is converted from Qwen/Qwen3.8-Flash-Next (BF16) with upstream Strata's tools.
  • tokenizer is the base model's tokenizer.
  • mmproj is not in this repository; it is downloaded from ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF (converted there from the base checkpoint).

Credits

Support (optional)

Sync-Strata and the Sync models are free. A donation page is at Ko-fi: a thank-you only, with nothing in return (no perks, no priority support, no feature promises), and it does not change the license terms.

License

Derived from Qwen3.8-Flash-Next: Qwen Community License 1.0 (LICENSE, Copyright (c) 2026 Qwen). Keep LICENSE with every copy. See the full text for the conditions on large commercial products and on Model-as-a-Service / AI work assistant businesses.