--- license: other license_name: swift-open-license-1.0 license_link: https://huggingface.co/ukisai/Swift-1.5-5bit-MLX/blob/main/LICENSE base_model: ukisai/Swift-1.5-Qwen3.8-27b base_model_relation: quantized library_name: mlx pipeline_tag: text-generation tags: - mlx - quantized - 5-bit - affine - qwen3_8 ---
UkisAI Website • Learn more • BF16 model • GGUF • GSQ-RCO GGUF • Evaluation • Enterprise licensing
# Swift 1.5 Qwen3.8-27B — 5-bit MLX **MLX affine 5-bit quantization, group size 64.** Swift 1.5 is UkisAI's reasoning-efficient Qwen3.8-27B derivative, focused on long-horizon, agentic and coding tasks. This export preserves the text, vision and MTP parameter tree; its supported generation interface is text-only with the included MLX-LM patches. Swift 1.5 uses **58.5% fewer thinking tokens** than base Qwen3.8-27B while scoring **0.35% higher**, for a **9.18× speed-up** on several tasks. > [!CAUTION] > Use only a complete snapshot whose files match `UPLOAD_MANIFEST.json`. > Historical build tests are not certification of an incomplete Hub snapshot. > Full independent Apple Silicon generation and quality evaluation remain **NOT_RUN**. > The 19.28 GB tensor payload must not be forced onto a 16 GiB Mac. ## Demo We gave base Qwen3.8-27B and Swift 1.5 27B the same prompt: > create a 3d little planet globe where I (player can walk around) and it has all these biomes to explore, the globe doesn't have to be too big, but still fun to go around. It's about a boy scout who is camping and goes around exploring. Try the game yourself here: [https://ukisai.com/swift-games/27b](https://ukisai.com/swift-games/27b) Base Qwen3.8-27B took 104.6 minutes to build its game. Swift 1.5 took 11.39 minutes. ## Source and quantization The recorded source is the complete customized Swift BF16 export at [`5ad04445d2686f525e9fbe5c077e6fa0c7df4200`](https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b/tree/5ad04445d2686f525e9fbe5c077e6fa0c7df4200), not base Qwen or another quantized model. The converter uses official MLX-LM commit `c69d1288440a0dc4e6401fc417098b07598dccd5` with the included architecture patch followed by the 5-bit extension. The original build report accounts for 1,199 source tensors, including 333 vision and 15 MTP tensors, and records 590 quantized modules plus 609 unquantized BF16 tensors after layout mapping. Its four shards contain 2,379 saved tensors and 19,281,804,384 bytes of tensor data. Independent recovery checks verified full SHA-256 hashes and header/index consistency of the original build. They did not repeat the complete source-value equality, finite-value or model-generation tests. The original tokenizer, template, configs, processors and quantized bytes are preserved. [QUANTIZATION_MANIFEST.json](QUANTIZATION_MANIFEST.json) and the existing `compatibility/` reports are historical build evidence. The current [upload manifest](UPLOAD_MANIFEST.json) identifies the intended complete file set. ## Evaluation See the [Swift BF16 source evaluation](https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b#evaluation) for source benchmarks and methodology. They were not rerun on this MLX export. No new broad accuracy, stability, long-context or BF16 quality-parity result is claimed. ## Validation and use **Both supplied MLX-LM patches are required, in order.** [USAGE.md](USAGE.md) contains the pinned install, full-file integrity check and generation example. The preserved source architecture is `Qwen3_5ForConditionalGeneration` / `qwen3_5`. The historical Linux build reports strict reload of 2,379 tensors, finite floating values, exact unquantized BF16 preservation, processor/tokenizer loading, and a short CPU generation returning `Hello from Swift.`. It also reports a real-weight vision encoder check and one explicit MTP step. These are historical results, not new independent inference results for the uploaded release. The CPU example promotes in-memory floating values to FP32 while retaining packed 5-bit UINT32 weights. A new synthetic macOS CPU/Metal diagnostic reproduces an MLX 0.32.2 CPU BF16 accumulation issue; it is not a Linux or full-model test. See [diagnostic](compatibility/macos-quantized-matmul-diagnostic.json) and [package checks](compatibility/package-checks.json). Full 27B Apple Silicon generation is unverified. Integrated image/video chat and speculative MTP generation are **not implemented** by the patch. Vision/MTP weights and component checks do not establish those end-to-end capabilities. Runtime/cache and OS memory must be budgeted in addition to the tensor payload. ## License and access Swift 1.5 derives from [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) (Copyright 2026 Alibaba Cloud, [Apache License 2.0](https://huggingface.co/ukisai/Swift-1.5-5bit-MLX/blob/main/LICENSE-APACHE-2.0)). UkisAI's adapted weights are licensed under the [Swift Open License v1.0](https://huggingface.co/ukisai/Swift-1.5-5bit-MLX/blob/main/LICENSE). See [NOTICE](https://huggingface.co/ukisai/Swift-1.5-5bit-MLX/blob/main/NOTICE) for attribution and change notices. Personal, research, educational, evaluation and commercial use are free for individuals and organizations with gross annual revenue, including affiliates, of up to US$1,000,000. Above that threshold, commercial use requires a separate Swift Enterprise License. Contact [UkisAI](https://ukisai.com/contact) for terms. Nothing in the Swift Open License limits the Apache 2.0 rights in Qwen3.8-27B itself. The accompanying Apple MLX-LM code has a separate upstream [MIT notice](compatibility/LICENSE-MLX-LM-MIT). ## Citation ```bibtex @misc{swift-1.5-qwen3.8-27b, title = {Swift 1.5 Qwen3.8-27B}, author = {UkisAI}, year = {2026}, url = {https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b} } ``` ## Acknowledgements We acknowledge the [NVIDIA Innovation Lab](https://www.nvidia.com/en-us/data-center/innovation-lab/), [Amazon Web Services](https://aws.amazon.com/), and [Google Cloud](https://cloud.google.com/) for compute credits and infrastructure support for Swift's development, training and evaluation.