[Website](https://ukisai.com)  • 
[Learn more](https://ukisai.com/products/swift)  • 
[GGUF](https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27B-GGUF)  • 
[GSQ-RCO GGUF](https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF)  • 
[Evaluation](#evaluation)  • 
[Enterprise licensing](#license-and-access)

Swift 1.5 Qwen3.8-27B

GGUF quantizations. Derived directly from Swift 1.5 with llama.cpp. Run a chosen quantization tier with a current llama.cpp-compatible runtime such as llama-server.

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.

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.

  Benchmark
  Final score
  Mean tokens
  Median tokens




  Qwen3.8
  Swift 1.5
  Qwen3.8
  Swift 1.5
  Reduction
  Reduction

General reasoning

GPQA-Diamond88.28%88.59%15,0148,717↓ 41.9%↓ 58.5%

C-Eval90.00%90.92%1,492819↓ 45.1%↓ 16.9%

IFBench73.53%72.07%8,0524,955↓ 38.5%↓ 47.3%

ERQA67.45%65.40%4,1371,906↓ 53.9%↓ 56.2%

Mathematics

AIME 202698.67%96.00%22,01413,203↓ 40.0%↓ 48.5%

HMMT November 202599.33%97.33%22,03214,957↓ 32.1%↓ 47.8%

Coding

LiveCodeBench v676.76%81.71%11,1848,448↓ 24.5%↓ 46.3%

Agent tasks

Terminal-Bench 2.169.21%72.13%52,26543,733↓ 16.3%↓ 0.1%

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.

Benchmark methodology and reproduction settings

Serving: BF16 · vLLM 0.27.1 · Qwen3 parser · context 262,144 · thinking xhigh.

Sampling: temperature 1.0 · top_p 0.95 · top_k 20 · min_p 0 · presence_penalty 0 · repetition_penalty 1.

Benchmarks: 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.

BenchmarkOutput cap

GPQA-Diamond100,000

C-Eval16,384

IFBench81,920

ERQA100,000

AIME 2026250,000

HMMT November 2025250,000

LiveCodeBench v632,768

Terminal-Bench 2.1Agent/task limits

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.

  Reasoning effort
  Qwen3.8
  Swift 1.5
  Mean thinking reduction

Xhigh88.28%88.59%↓ 41.9%

Medium84.14%82.22%↓ 24.8%

Low84.04%84.85%↓ 28.7%

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.

Benchmark / Swift 1.5 quantization
Qwen base  

accuracy Swift 1.5 quant
accuracy Mean token reduction Median token reduction

GPQA-Diamond
AWQ INT486.36%88.38%↓ 51.5%↓ 64.4%

GPQA-Diamond
AutoRound INT486.36%89.39%↓ 50.5%↓ 57.8%

GPQA-Diamond
AWQ + GPTQ INT486.36%90.91%↓ 45.8%↓ 64.4%

IFBench
AWQ INT472.00%72.00%↓ 36.9%↓ 49.3%

IFBench
AutoRound INT472.00%69.33%↓ 29.3%↓ 39.2%

IFBench
AWQ + GPTQ INT472.00%70.00%↓ 31.8%↓ 52.7%

AIME 2026
AWQ INT470.00%86.67%↓ 29.2%↓ 36.2%

AIME 2026
AutoRound INT476.67%83.33%↓ 17.7%↓ 32.4%

AIME 2026
AWQ + GPTQ INT476.67%83.33%↓ 22.4%↓ 34.0%

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.

Quantized evaluation settings and BF16 reference

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.

GGUF quantizations

File
Size
KLD wikitext @512
KLD wikitext @32k
99% KLD @32k
Top-p @32k

Q8_029.0 GB0.00080.00060.00598.85%

Q6_K_L25.0 GB0.00150.00140.01098.16%

Q6_K23.9 GB0.00180.00160.01498.30%

Q6_K_S22.9 GB0.00200.00160.01498.24%

Q5_K_M20.9 GB0.00520.00610.05096.92%

Q5_K_S19.6 GB0.00600.00690.05896.91%

Q4_K_L18.8 GB0.01060.01030.10595.79%

Q4_K_M17.4 GB0.01370.01340.16395.03%

IQ4_NL17.4 GB0.01520.01400.17595.39%

Q4_K_S16.4 GB0.01640.01540.17594.84%

IQ4_XS15.5 GB0.01790.01730.18794.96%

IQ3_M14.9 GB0.04100.03800.40991.83%

Q3_K_L14.1 GB0.04420.04100.41591.63%

Q3_K_M13.4 GB0.05700.05620.61490.30%

IQ3_XS12.8 GB0.05830.08851.13088.54%

Q3_K_S12.7 GB0.06480.06580.71289.53%

IQ3_XXS12.3 GB0.07420.08440.99688.70%

Q2_K10.8 GB0.16550.15461.72884.00%

IQ2_M10.5 GB0.15230.14931.56884.17%

IQ2_S9.7 GB0.20950.25893.02480.58%

IQ2_XS9.1 GB0.24330.26222.95179.99%

IQ2_XXS8.9 GB0.28660.27692.90278.48%

Mean KL divergence against the Swift 1.5 BF16 source, lower is better. wikitext @512 is wikitext-2 test, 100 windows of 512 tokens. wikitext @32k is wikitext-2 train, 16 windows of 32,768 tokens, scoring only the last 512 tokens of each window, so every scored token sees at least 32k tokens of context. 99% KLD is the 99th-percentile divergence on the same 32k run. Top-p is top-token agreement with BF16 on the 32k run.

Long context costs very little on this release: for every tier from Q8_0 through Q4_K_S, the 32k mean is within 10% of the 512-token value (Q5_K_M and Q5_K_S rise about 15%), and Q4_K_M keeps the same top token as BF16 on 95% of positions at 32k. The pick below follows the 99th-percentile tail at 32k: 0.163 for Q4_K_M, 0.050 for Q5_K_M, 0.014 for Q6_K, 0.005 for Q8_0.

Use case
Pick

24 GB cards, everyday useQ4_K_M

Long agentic runs, strict tool-call formattingQ6_K or higher

Maximum fidelityQ8_0

Recipe

All 22 tiers were built with llama.cpp commit 6f41ac5 from a BF16 conversion of the published Swift 1.5 safetensors. They reuse the importance matrix and the per-tensor type layouts (--tensor-type-file) that bartowski computed for Swift 1.0 with his quantization-config; Swift 1.5 has the same architecture and tensor shapes. Every file was checked against BF16 on the harness above.

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.