SJF Swift1.5 NInfer GSQ-RCO IQ3_S MTP

This is an unofficial format conversion, not a newly trained model or an official release from UkisAI, WaveCut, or Alibaba. It converts UkisAI's Swift-1.5-Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf to the WaveCut NInfer v3 container while retaining the source GGUF tensor quantization and MTP head. SJF identifies this conversion and test release.

Sources and credit

Component	Source	Role
Base model	Qwen3.8-27B, Alibaba Cloud	Base weights and architecture
Swift 1.5	UkisAI Swift 1.5 GSQ-RCO GGUF	Adapted weights and GGUF release
GSQ-related research	ISTA-DASLab, as credited by the source release	Quantization work
NInfer runtime	WaveCut/ninfer-all	Runtime and v3 container format
SJF release	Conversion, Windows compatibility patch, and local validation	This distribution

The weights remain subject to the Swift Open License v1.0, including its commercial-use terms. The Qwen base model's Apache 2.0 license and the source NOTICE are included. The NInfer code and the separate Windows patch are under Apache 2.0. Read the licenses before redistribution or commercial use.

Commercial use and authorization

Redistribution. This conversion may be reproduced and redistributed under Section 4 of the Swift Open License v1.0. Every copy must carry that license, the attribution notices in NOTICE, and the Qwen base-model license (Apache 2.0).

Commercial-use limitation. Under Section 5 of that license, an entity — together with every entity that controls it, is controlled by it, or is under common control with it — whose gross revenue for the most recently completed fiscal year is US$1,000,000 or more is not licensed for commercial use of these weights.

Enterprise license. An entity above that threshold must obtain a separate written Swift Enterprise License from the Licensor, UkisAI. Requests: https://ukisai.com/contact. That license can be granted by UkisAI only. This release and its maintainer are not agents of UkisAI: they cannot grant, sublicense, or waive any commercial right in the upstream weights, and they cannot receive or approve such a request.

Below the threshold. Commercial use by entities below the threshold is licensed under the Swift Open License v1.0. Non-commercial and research use is not subject to the threshold.

This release's own content. The conversion tooling, assemble_model.py, the Windows compatibility patch, and this documentation are licensed under the Apache License 2.0.

Statement of origin. This release is an unofficial format conversion of UkisAI's Swift 1.5 weights. It is not affiliated with, sponsored by, or endorsed by UkisAI, WaveCut, or Alibaba Cloud.

Download and verify

The model is distributed as ordered .partNN assets in the GitHub Release because it exceeds GitHub's per-asset limit. Download all parts into one directory, then run:

python .\assemble_model.py --parts-dir . --output .\SJF-Swift1.5-NInfer-GSQ-RCO-IQ3_S-MTP.ninfer

The script checks every part's SHA-256, concatenates them in order, and checks the final model hash. It will not overwrite an existing output. The script and SHA256SUMS.json are in this repository and attached to the Release.

File	Bytes	SHA-256
Source GGUF	12,120,016,896	9aecf1cd41b2cb2f32a74e0d889e33855ebef43b26f43b43feb5720239e677e5
Converted NInfer v3	12,494,820,096	152307059a0d9a7e31b9f9db63fd568d29160863c128bd92869d33d372ab0908

The source GGUF hash matches the upstream checksum file. This repository does not redistribute the GGUF.

Build and compatibility

• Conversion/runtime source: WaveCut ninfer-all commit 8ed38f69670e7dfe2b5db18b1081a3b3085c4305 plus the included ninfer-windows-compat.patch.
• The patch records five Windows / RTX 5090 build compatibility changes. It does not include a copy of the upstream runtime or third-party binary dependencies.
• Model format: NInfer v3. Tested with the above runtime on Windows and an RTX 5090. Other NInfer forks, Linux builds, and other GPUs have not been validated with this file.
• The converted model is approximately 12.5 GB on disk. That is not the complete GPU memory requirement.

See CONVERSION.md for the conversion record and BENCHMARK.md for test setup and results.

Measured behavior

Three different, independently checked xhigh tasks were run on Windows / RTX 5090 with single concurrency, temperature 0, MTP 3, 262,144 configured context capacity, and 98,304 maximum output setting. The converted Swift + NInfer model answered all three correctly and averaged 230.6 output tokens/s and 81.6 seconds per task measured from request to completed response. Average output was 18,939 tokens per task. The complete per-task comparison with source Swift GGUF + llama.cpp and ordinary GSQ + NInfer is in BENCHMARK.md.

Long-form writing caveat: In two later DeepSeek Harness medium reasoning turns, pure decode speed was 149.4 and 151.7 tok/s, and both responses missed a 5,000-Chinese-character request. One response incorrectly claimed it was about 5,200 characters. See BENCHMARK.md for the actual counts and request conditions. The xhigh task mean is not a general writing-speed or instruction-following guarantee.

The 256K setting was a capacity configuration, not a full 256K input test. The 96K output setting was an upper limit; no test generated 96K tokens continuously. The three-task result does not establish a fixed 400 tokens/s speed, 12 GB total VRAM use, or production stability.

Changes from upstream

1. Converted UkisAI's named GGUF into the NInfer v3 container without retraining or re-quantizing the source tensors.
2. Added NInfer container resources and indices required by the WaveCut runtime.
3. Applied five Windows compatibility changes to the runtime build, provided separately as ninfer-windows-compat.patch.
4. Published conversion checksums, assembly utility, and local test results under this SJF release name.

Swift, Qwen, WaveCut, and other names identify their respective source projects; this release is not endorsed by those projects.