Swift-1.5-Qwen3.8-27b-NVFP4

ukisai/Swift-1.5-Qwen3.8-27b-NVFP4

Swift-1.5-Qwen3.8-27b-NVFP4 is a image-text-to-text model from ukisai. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for vllm. The card lists the license as other.

<div align="center" <a href="https://ukisai.com"<img src="ukisai-banner.png" alt="UkisAI" style="width:100%;max-width:100%;height:auto;display:block;margin-bottom:0.6em;" /</a <div style="display:flex;justify-content:…

Downloads · 30 days

5.5K

42% of all-time downloads

All-time downloads

13.2K

Public

Parameters

18.2B

21.9 GB on disk

Likes

Trending 3

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors21.9 GB · 100%

Parameter types

How the weights are stored.

U89.2B · 49%

Try a prompt

Base models

ukisai/Swift-1.5-Qwen3.8-27b

Spaces

ukisai/Swift-27B

Model card

From the Hugging Face model README

<div align="center"> <a href="https://ukisai.com"><img src="ukisai-banner.png" alt="UkisAI" style="width:100%;max-width:100%;height:auto;display:block;margin-bottom:0.6em;" /></a> <div style="display:flex;justify-content:center;gap:0.6em;margin-bottom:1em;"> <a href="https://ukisai.com"><strong>Website</strong></a>  •  <a href="https://ukisai.com/products/swift"><strong>Learn more</strong></a>  •  <a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27B-GGUF"><strong>GGUF</strong></a>  •  <a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF"><strong>GSQ-RCO GGUF</strong></a>  •  <a href="#evaluation"><strong>Evaluation</strong></a>  •  <a href="#license-and-access"><strong>Enterprise licensing</strong></a> </div> </div>

Swift 1.5 Qwen3.8-27B

NVIDIA NVFP4/FP8 mixed precision. Derived directly from Swift 1.5 with NVIDIA ModelOpt. Run it with a ModelOpt-compatible runtime; native NVFP4 execution requires compatible NVIDIA Blackwell support.

Swift 1.5 Qwen3.8-27B is UkisAI's reasoning-efficient derivative of Qwen3.8-27B. It uses 58.5% fewer thinking tokens while scoring 0.35% higher than the base, for a 9.18× speed-up on several tasks.

Conversion provenance is pinned to Swift 1.5 source revision 71e23232d65779d3087ba31dd383cddba6ba48df, as recorded in UPLOAD_MANIFEST.json. The source repository was subsequently completed with all 18 BF16 shards and runtime assets at revision 5ad04445d2686f525e9fbe5c077e6fa0c7df4200. The complete-revision link is provided for reproducible downloading and does not change the actual conversion provenance.

Swift 1.5 is a direct upgrade from Swift 1.0, our model with 350k+ downloads, delivering stronger overall performance than both base and Swift 1.0 in various tasks, especially coding and agentic, while using fewer thinking tokens. We accomplished that by scaling up the post-training (RL and OPD) from the previous version.

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

Base Qwen3.8-27B took 104.6 minutes to build its game. Swift 1.5 took 11.39 minutes.

Training approach

We made Swift efficient by figuring out which tokens were linked to pathological overthinking and penalizing them without "attacking" the reasoning length directly then regained the accuracy with RL and OPD, leading to "compressed" token usage while maintaining accuracy. Swift 1.5 was made from Swift 1.0, on whom we scaled up the post-training methods that previously improved Swift1.0 model performance, this time with the main focus on long-horizon, agentic, and coding tasks, as seen in the LiveCodeBench and Terminal Bench 2.1 improvements. Our training data is viewable here: https://huggingface.co/datasets/ukisai/Qwen3.8-27B-multi-turn-agent-sft albeit it is not used out of the box, but rather re-sampled, turned into proper RL environments etc.

Evaluation

The external results below compare Qwen3.8-27B, the foundation base model, and Swift 1.5. Both models use the same saved evaluation protocols, and all scores are reported as final aggregate percentages.

Scores are final five-repeat aggregates under matched evaluation protocols. Mean-token columns report reasoning tokens per trial; Terminal-Bench sums reasoning across agent calls.

<details> <summary><strong>Benchmark methodology and reproduction settings</strong></summary> <p style="font-size:13px;line-height:1.5;margin:8px 0;"><strong>Serving:</strong> BF16 · vLLM 0.27.1 · Qwen3 parser · context 262,144 · thinking xhigh.<br> <strong>Sampling:</strong> temperature 1.0 · top_p 0.95 · top_k 20 · min_p 0 · presence_penalty 0 · repetition_penalty 1.<br> <strong>Benchmarks:</strong> averages over five seeds (0–4) per model; five trials per task for Terminal-Bench, base and Swift 1.5 served at context 131,072 on the same Harbor build.</p> <table style="display:table;width:100%;border-collapse:collapse;font-size:13px;line-height:1.3;margin:8px 0;"> <thead><tr><th style="padding:4px 8px;text-align:left;">Benchmark</th><th style="padding:4px 8px;text-align:right;">Output cap</th></tr></thead> <tbody> <tr><td style="padding:3px 8px;">GPQA-Diamond</td><td style="padding:3px 8px;text-align:right;">100,000</td></tr> <tr><td style="padding:3px 8px;">C-Eval</td><td style="padding:3px 8px;text-align:right;">16,384</td></tr> <tr><td style="padding:3px 8px;">IFBench</td><td style="padding:3px 8px;text-align:right;">81,920</td></tr> <tr><td style="padding:3px 8px;">ERQA</td><td style="padding:3px 8px;text-align:right;">100,000</td></tr> <tr><td style="padding:3px 8px;">AIME 2026</td><td style="padding:3px 8px;text-align:right;">250,000</td></tr> <tr><td style="padding:3px 8px;">HMMT November 2025</td><td style="padding:3px 8px;text-align:right;">250,000</td></tr> <tr><td style="padding:3px 8px;">LiveCodeBench v6</td><td style="padding:3px 8px;text-align:right;">32,768</td></tr> <tr><td style="padding:3px 8px;">Terminal-Bench 2.1</td><td style="padding:3px 8px;text-align:right;">Agent/task limits</td></tr> </tbody> </table> </details>

Efficiency across reasoning efforts

Qwen3.8's reasoning_effort setting lets users choose how much the model thinks. For Swift 1.5 to be useful across these settings, it needs to reduce thinking while keeping accuracy close to the base. We therefore tested xhigh, medium, and low: thinking-token savings persist at every level.

<table class="swift15-table swift15-compact"> <thead> <tr> <th class="benchmark-heading" style="width:34%;text-align:left;padding-left:18px;">Reasoning effort</th> <th class="score-heading" style="width:22%;">Qwen3.8</th> <th class="score-heading" style="width:22%;">Swift 1.5</th> <th class="tokens-heading" style="width:22%;">Mean thinking reduction</th> </tr> </thead> <tbody> <tr><td class="model">Xhigh</td><td>88.28%</td><td class="swift15"><strong>88.59%</strong></td><td class="change">↓ 41.9%</td></tr> <tr><td class="model">Medium</td><td><strong>84.14%</strong></td><td class="swift15">82.22%</td><td class="change">↓ 24.8%</td></tr> <tr><td class="model">Low</td><td>84.04%</td><td class="swift15"><strong>84.85%</strong></td><td class="change">↓ 28.7%</td></tr> </tbody> </table>

At low, Swift 1.5 scores above the base while using about 29% fewer thinking tokens.

Quantized Swift 1.5 models

Format	Repository	Runtime
GGUF	Swift-1.5-Qwen3.8-27B-GGUF	llama.cpp
GSQ-RCO GGUF (compact 2–3 bit)	Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF	llama.cpp
AWQ INT4 (W4A16)	Swift-1.5-Qwen3.8-27b-W4A16-AWQ	vLLM (`compressed-tensors`)
AutoRound INT4 (W4A16)	Swift-1.5-Qwen3.8-27b-W4A16-AutoRound	vLLM (`auto-round`)
AWQ + GPTQ INT4 (W4A16)	Swift-1.5-Qwen3.8-27b-INT4	vLLM (`compressed-tensors`)
NVFP4	Swift-1.5-Qwen3.8-27b-NVFP4	NVIDIA Blackwell
AMD Quark FP8 (W8A8)	Swift-1.5-Qwen3.8-27b-Quark-FP8-dynamic-AMD	AMD Quark
MLX 5-bit	Swift-1.5-5bit-MLX	Apple MLX
MLX 4-bit	Swift-1.5-4bit-MLX	Apple MLX
MLX 3-bit (text only)	Swift-1.5-3bit-MLX-TextOnly	Apple MLX

These results evaluate the merged Swift 1.5 checkpoint and three INT4 exports on GPQA-Diamond (198 questions), IFBench (300 prompts), and AIME 2026 (30 problems). Each model completed the full datasets with one sample per prompt, seed 0, and zero request errors. This is a single-seed evaluation, separate from the five-repeat BF16 release results above.

The Qwen-base columns use the saved seed/sample 0 runs. Quantization recipes and serving settings differ from the new Swift 1.5 runs, so these are reference comparisons rather than a controlled measurement of the Swift adaptation. Token reductions below are recomputed from those same reference samples.

<table class="swift15-table" style="display:table;width:100%;table-layout:fixed;"> <thead><tr> <th class="benchmark-heading" style="width:32%;text-align:left;padding-left:18px;">Benchmark / Swift 1.5 quantization</th> <th class="score-heading" style="width:16%;">Qwen base<br>accuracy</th> <th class="score-heading" style="width:16%;">Swift 1.5 quant<br>accuracy</th> <th class="tokens-heading" style="width:18%;">Mean token reduction</th> <th class="median-heading" style="width:18%;">Median token reduction</th> </tr></thead> <tbody> <tr><td class="model">GPQA-Diamond<br><span class="detail"><a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b-W4A16-AWQ">AWQ INT4</a></span></td><td>86.36%</td><td class="swift15">88.38%</td><td class="change">↓ 51.5%</td><td class="change">↓ 64.4%</td></tr> <tr><td class="model">GPQA-Diamond<br><span class="detail"><a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b-W4A16-AutoRound">AutoRound INT4</a></span></td><td>86.36%</td><td class="swift15">89.39%</td><td class="change">↓ 50.5%</td><td class="change">↓ 57.8%</td></tr> <tr><td class="model">GPQA-Diamond<br><span class="detail"><a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b-INT4">AWQ + GPTQ INT4</a></span></td><td>86.36%</td><td class="swift15">90.91%</td><td class="change">↓ 45.8%</td><td class="change">↓ 64.4%</td></tr> <tr><td class="model">IFBench<br><span class="detail"><a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b-W4A16-AWQ">AWQ INT4</a></span></td><td>72.00%</td><td class="swift15">72.00%</td><td class="change">↓ 36.9%</td><td class="change">↓ 49.3%</td></tr> <tr><td class="model">IFBench<br><span class="detail"><a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b-W4A16-AutoRound">AutoRound INT4</a></span></td><td>72.00%</td><td class="swift15">69.33%</td><td class="change">↓ 29.3%</td><td class="change">↓ 39.2%</td></tr> <tr><td class="model">IFBench<br><span class="detail"><a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b-INT4">AWQ + GPTQ INT4</a></span></td><td>72.00%</td><td class="swift15">70.00%</td><td class="change">↓ 31.8%</td><td class="change">↓ 52.7%</td></tr> <tr><td class="model">AIME 2026<br><span class="detail"><a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b-W4A16-AWQ">AWQ INT4</a></span></td><td>70.00%</td><td class="swift15">86.67%</td><td class="change">↓ 29.2%</td><td class="change">↓ 36.2%</td></tr> <tr><td class="model">AIME 2026<br><span class="detail"><a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b-W4A16-AutoRound">AutoRound INT4</a></span></td><td>76.67%</td><td class="swift15">83.33%</td><td class="change">↓ 17.7%</td><td class="change">↓ 32.4%</td></tr> <tr><td class="model">AIME 2026<br><span class="detail"><a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b-INT4">AWQ + GPTQ INT4</a></span></td><td>76.67%</td><td class="swift15">83.33%</td><td class="change">↓ 22.4%</td><td class="change">↓ 34.0%</td></tr> </tbody> </table>

AIME scoring: truncated responses count as incorrect for both columns.

The AMD Quark INT4 and FP8 exports have separate sanity evaluations; completed results on these three reasoning benchmarks are not available for them.

<details> <summary><strong>Quantized evaluation settings and BF16 reference</strong></summary>

Serving: vLLM 0.29.0, tensor parallelism 1, eager execution, BF16 activations, context 131,072, template-default thinking without an effort override. The AWQ + GPTQ export uses FP8 KV cache; BF16, AWQ, and AutoRound use auto KV dtype. Sampling: temperature 1, top-p 0.95, top-k 20, min-p 0, presence penalty 0, repetition penalty 1, seed 0. Output caps: GPQA 100,000, IFBench 81,920, AIME 32,768. IFBench uses official strict prompt-level scoring.

GPQA token counts cover re-tokenized reasoning; IFBench and AIME count the full generated response. Statistics include all responses, including truncations; medians use the midpoint of the two central values when the sample count is even.

Saved Qwen references: W4A16 for GPQA and IFBench; Qwen AWQ for the AWQ AIME row; Qwen W4A16 for the AutoRound and AWQ + GPTQ AIME rows. The latter is a W4A16 reference for AutoRound, not an AutoRound base run. The new runs do not reproduce the original software stack.

The fresh Swift 1.5 BF16 reference and all quantized exports scored as follows under this single-seed protocol:

Model	GPQA-Diamond	IFBench strict	AIME 2026
Swift 1.5 BF16	91.41%	72.00%	86.67%
AWQ INT4	88.38%	72.00%	86.67%
AutoRound INT4	89.39%	69.33%	83.33%
AWQ + GPTQ INT4	90.91%	70.00%	83.33%

Truncation counts are recorded in the linked evaluation data. These single-seed results do not establish quality parity or replace the broader multi-seed evaluation.

Verified counts, token statistics, settings, and evidence hashes.

</details>

How to run

This is a unified Hugging Face ModelOpt mixed NVFP4/FP8 checkpoint. NVIDIA's current Qwen3.8 NVFP4 reference uses the vllm/vllm-openai:nightly runtime on Blackwell hardware. Use a current vLLM nightly on a compatible NVIDIA Blackwell GPU and start with a context length that fits the available memory:

vllm serve ukisai/Swift-1.5-Qwen3.8-27b-NVFP4 \
  --revision 332d277a083ca395aa2f822842b82f5f2f3913a2 \
  --dtype bfloat16 \
  --kv-cache-dtype auto \
  --mamba-ssm-cache-dtype float32 \
  --tensor-parallel-size 1 \
  --max-model-len 32768 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --port 8000

KV cache and recurrent state

Weight quantization and cache precision are separate. This export declares kv_cache_quant_algo: null in hf_quant_config.json, has no kv_cache_scheme in config.json, and contains no checkpoint Q/K/V cache-scale tensors. For the command above, --dtype bfloat16 --kv-cache-dtype auto is the intended BF16 attention-cache baseline; it does not request FP8 or NVFP4 KV cache. The 48 linear-attention layers use separate recurrent state. Preserve the checkpoint's mamba_ssm_dtype: float32; the explicit SSM flag above makes this requirement visible instead of relying on runtime-version-dependent defaults.

Do not enable FP8/NVFP4 KV cache merely because the weights are quantized. Forcing per-tensor FP8 without calibrated checkpoint scales can use unit scales and change output quality. Such a profile requires its own compatible backend, scale validation, and comparison against the BF16-cache baseline. Projection weight/input scales are not KV-cache scales. Old runtime advice such as --calculate-kv-scales must be checked against the exact installed version; it is absent from the vLLM revision inspected on 2026-09-23.

Pin the serving container digest or vLLM commit as well as the model revision. Source-level configuration tests do not establish a tested serving image: resolved GPU cache dtypes, cache reuse, long-context stability and full NVFP4 generation remain independently unverified. Check those in startup logs and real prefill/decode tests before claiming runtime success.

Increase --max-model-len or tensor parallelism only after checking available GPU memory. The checkpoint inventory, serialization and checksums were validated, but full end-to-end generation from this Swift 1.5 NVFP4 export was not independently tested. The evaluation tables above describe the linked BF16/INT4 evaluation data; they are not independent NVFP4 benchmark results.

License and access

Swift 1.5 is a derivative of Qwen3.8-27B (Copyright 2026 Alibaba Cloud, Apache License 2.0). UkisAI's contribution, including the adapted weights, is licensed under the Swift Open License v1.0. See NOTICE for the change notice and attribution details.

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 for terms.

Nothing in the Swift Open License limits rights in Qwen3.8-27B itself under Apache 2.0.

Citation

@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, Amazon Web Services, and Google Cloud for providing compute credits and infrastructure support for Swift's development, training, and evaluation.

Task: Image-Text-to-Text
Library: vllm
Type: qwen3_5
License: other
Created: Sep 21, 2026
Updated: Sep 24, 2026