Minachist/Swift1.5-Qwen3.8-Flash-Next-INT4-Mixed-AutoRound

Swift 1.5 Qwen3.8-Flash-Next, INT4/INT6 mixed (AutoRound), for vLLM on two, three, four 24 GB GPUs or two 32 GB GPUs

This is an unofficial quantization. It was made by Minachist. It is not made, reviewed, endorsed or supported by UkisAI or Qwen. The weights come from Swift 1.5 Qwen3.8-Flash-Next by UkisAI, which is a fine-tune of Qwen3.8-Flash-Next by Qwen. The names "UkisAI" and "Swift" are used here only to say where the weights come from.

Same recipe as the new-main branch of Minachist/Qwen3.8-Flash-Next-INT4-Mixed-AutoRound: the calibration data is written by the model that is quantized, here Swift 1.5. The model runs in vLLM on three RTX 3090 (24 GB) with the patches from that repository.

License

Two licenses apply to these weights. Both are in this repository, unchanged.

File	License	Who it is from
LICENSE	Swift Open License v1.0	UkisAI. Covers UkisAI's contribution, which is how Swift 1.5 differs from Qwen3.8-Flash-Next.
LICENSE-QWEN	Qwen Community License 1.0	Qwen. Covers Qwen3.8-Flash-Next.

The licenses still apply after quantization. This repository adds no license restrictions of its own.

• Swift commercial use: section 5 sets a US$1,000,000 gross-revenue threshold, including entities under the control relationships defined in section 1, measured over the most recently completed fiscal year. Commercial use by a legal entity that exceeds the threshold requires a separate written license from UkisAI. The qualified non-profit exception is for non-commercial or research purposes. Read sections 1 and 5, including their exact boundary wording.
• Qwen commercial use: if you or an affiliate operate a Model as a Service or AI Work Assistant business, Qwen requires a separate license before commercial use, subject to the internal-use exception in condition 2. A commercial product or service exceeding 100,000,000 monthly active users or US$20,000,000 monthly revenue must prominently display the respective model name in its UI under condition 1.
• Redistribution: preserve the applicable licenses and notices, and mark modified files as required by section 4. Under section 12, noncompliance terminates the Swift license and requires stopping use and deleting copies containing the Swift contribution; rights in the base model remain separate.

The redistributor notice from the appendix of LICENSE is reproduced here:

Copyright 2026 UkisAI. Swift Contribution licensed under the Swift Open License v1.0 (https://huggingface.co/ukisai/Swift-Qwen3.8-Flash-Next/blob/main/LICENSE). Derivative of Qwen3.8-Flash-Next, Copyright (c) 2026 Qwen, Qwen Community License 1.0.

Changes from Swift 1.5 Qwen3.8-Flash-Next

NOTICE has the same list, with UkisAI's own NOTICE in front of it.

• model-*.safetensors, model_extra_tensors.safetensors, model.safetensors.index.json: the weights are quantized (see "Quantization") and written to new files.
• config.json: a quantization_config block was added, and the exporter rewrote a few fields. quantization_config.json and processor_config.json were added by the export tool, and it added a processor_class entry to tokenizer_config.json.
• README.md is new.
• File-local modification notices were added by Minachist on 2026-10-09: in the __metadata__ of all 48 safetensors files, in the index's metadata, and as _modification_notice in the exported configuration JSON files (including the embedded quantization configuration). This last step changed no tensor values or inference settings. NOTICE was extended to document it.

These files are identical copies of the Swift 1.5 files. LICENSE, LICENSE-QWEN, generation_config.json, chat_template.jinja, vocab.json, merges.txt, tokenizer.json, preprocessor_config.json, video_preprocessor_config.json

UkisAI's model card, banner, demo video or benchmark files are not included. Their benchmark numbers are for the BF16 model, and it's not for this quantization.

Quantization

Everything below uses compressed-tensors, pack-quantized, and symmetric group quantization.

Group	Scheme	What it covers
A	INT4, gs128	MoE routed experts (512 per layer, top-10). 92% of the body
B	INT6, gs64	linear_attn in and out projections, QSA q/k/v/o_proj, shared expert
C	INT8, gs64	hyper-connection low-rank mixers
D	INT8, gs128	lm_head, embed_tokens, PLE key_proj and value_proj, indexer index_qk_proj
E	BF16	router, vision tower, PLE n-gram table, MTP module (same bits as Swift 1.5)

Running

The model has the same architecture, tensor names and shapes as Qwen3.8-Flash-Next, so it runs with the same vLLM patches. They are not in this repository. Use the build with patches from Minachist/Qwen3.8-Flash-Next-INT4-Mixed-AutoRound (vllm-patch/, which also has the requirements, the settings and the other builds), and point MODEL_DIR at this repository. Only 3x24GB was tested with this model.

Notes

• The vLLM patches are a separate work with their own licenses (Apache-2.0 and MIT, see vllm-patch/NOTICE in the repository above). They are not covered by the licenses of the model.

• In order to publish this, I did:

  1. Read the licenses in full. Swift 1.5 carries the Swift Open License v1.0 alongside the Qwen Community License 1.0. I reviewed their definitions, redistribution conditions, commercial restrictions, trademark provisions and termination language to determine what applied to this quantization.
  2. Provide both licenses to recipients. Sections 4(a) and 4(e) require copies of the Swift license and the base-model license. I included LICENSE and LICENSE-QWEN unchanged. Section 4(e) expressly includes quantized weights that incorporate the base model.
  3. Preserve the applicable upstream notices and distinguish my additions. Sections 4(c) and 4(d) require retention of the applicable copyright, patent, trademark and attribution notices. I retained the complete upstream NOTICE unchanged at the beginning of this repository's NOTICE, then appended a separate account of my modifications. My additions do not modify either license.
  4. Identify the files I changed. Section 4(b) requires modified files to carry prominent notices stating that they were changed. I documented the quantized and re-sharded weights, the weight index, the added quantization and processor configurations, and the fields rewritten by the exporter, including its addition to tokenizer_config.json. The README and NOTICE describe these changes.
  5. Make the modified files themselves carry notices. To address the wording of section 4(b), I added modification notices to the metadata of all 48 safetensors files, the weight index and the exported configuration JSON files. The embedded quantization configuration and its standalone file received matching notices.
  6. Keep my use of upstream names within the trademark provision. Section 8 grants no general permission to use UkisAI's names or trademarks, while allowing reasonable and customary use to describe the work's origin and reproduce notices. I used the names for attribution, identified this as an unofficial quantization, and wrote an independent model card.

• These requirements also come with section 12: failure to comply with any condition causes automatic, immediate termination, followed by an obligation to stop using and delete copies containing the Swift contribution. It specifies no cure period or reinstatement procedure.

Acknowledgements

• vLLM (Apache-2.0), for the inference engine, GPU kernels, and serving infrastructure used to generate calibration data and run this quantized model.

• AutoRound (Apache-2.0), for the quantization toolkit, weight optimization, and export support used to produce this checkpoint.

• Arch Linux, for the operating system, packaging ecosystem, and documentation behind the environment used for this work.

• aikitoria (MIT/GPL-2.0), for the patched GPU kernel modules enabling peer-to-peer transfers between consumer NVIDIA GPUs, used on the RTX 3090 system behind this work.

• tinygrad (MIT/GPL-2.0), for the original peer-to-peer patch that these modules build on, and NVIDIA (MIT/GPL-2.0), for publishing the open GPU kernel modules themselves.

• Qwen (Qwen Community License 1.0), for developing Qwen3.8-Flash-Next and publishing the base weights and architecture on which this work builds.

Thank you to everyone above who continues to make their work available under clear, straightforward, concise, permissive and easy to understand licensing terms, allowing others to use, modify, and share it without having to worry about inadvertently violating them.