ukisai/Swift-1.5-Ornith-9B-GGUF

Website  •  Learn more  •  Safetensors  •  GGUF files  •  Evaluation  •  License

Swift 1.5 Ornith 9B GGUF

GGUF. BF16 and quantized GGUF files of Swift 1.5 Ornith 9B made with llama.cpp. Every language-model file includes the MTP head; the vision encoder is in the separate mmproj file.

Swift 1.5 Ornith 9B is UkisAI's reasoning-efficient derivative of Ornith-1.5-9B. On GPQA-Diamond it uses 25.2% fewer thinking tokens (33.1% fewer on the median) and scores 1.52 pp higher than the base. Across our benchmarks it uses 6–25% fewer thinking tokens on the mean and 12–34% fewer on the median, while staying within 2.3 pp of the base on AIME, IFBench and LiveCodeBench.

This is an experimental Swift release. See Research status and limitations.

Training approach

We built Swift by identifying reasoning-marker tokens that, in our analysis, trigger overthinking in the model's reasoning rollouts. We then fine-tuned the model by penalizing usage of those tokens while it reasons.

Swift produces shorter reasoning traces while keeping accuracy close to the base model. In our testing, Swift 1.5 also hits the output cap less often than the base on every benchmark below.

The vision encoder, the multi-token prediction (MTP) layer, the token embeddings and the output head are unchanged from Ornith-1.5-9B.

Evaluation

All scores compare Ornith-1.5-9B with Swift 1.5 Ornith 9B, both in BF16 with thinking on. Both models used the same harness, sampling, seeds and output cap on every row. Token columns report thinking tokens per attempt; the visible answer is not counted.

Benchmark	Score		Mean tokens			Median tokens
	Base	Swift 1.5	Base	Swift 1.5	Reduction	Reduction
General reasoning
GPQA-Diamond	83.00%	84.51%	15,760	11,782	↓ 25.2%	↓ 33.1%
IFBench	42.60%	40.47%	13,409	10,790	↓ 19.5%	↓ 24.7%
Mathematics
AIME 2026	92.67%	90.67%	37,872	35,460	↓ 6.4%	↓ 29.2%
AIME 2024	94.00%	93.33%	20,776	17,266	↓ 16.9%	↓ 12.2%
Coding
LiveCodeBench v6	59.20%	56.91%	41,057	34,343	↓ 16.4%	↓ 33.5%

Swift 1.5 is ahead of base on GPQA-Diamond (+1.52 pp). It trails base on AIME 2024 (−0.67 pp), AIME 2026 (−2.00 pp), IFBench (−2.13 pp) and LiveCodeBench v6 (−2.29 pp).

How to reproduce

Serving: BF16 · tensor parallelism 1 · context 262,144 · thinking on.
Sampling: temperature 1.0 · top_p 0.95 · top_k 20 · min_p 0 · presence_penalty 0.
Scoring: every attempt counts. Request errors, answers cut off at the output cap and empty answers count as wrong.
GPQA-Diamond: 198 questions × 3 seeds. The answer letter comes from the final answer, not from the thinking.
IFBench: 300 prompts × 5 seeds, official prompt-level loose score.
AIME 2026: 30 problems × 5 seeds, reasoning effort xhigh, scorer from UkisAI Benchmark_configs.
AIME 2024: 30 problems from MathArena × 5 seeds.
LiveCodeBench v6: the 175 problems that release v6 added, official pass@1 averaged over 5 seeds, 10-second limit per test.

Benchmark	Output cap	Cap hits, base	Cap hits, Swift 1.5
GPQA-Diamond (594 attempts)	81,920	48	24
IFBench (1,500 attempts)	81,920	86	69
AIME 2026 (150 attempts)	250,000	2	0
AIME 2024 (150 attempts)	250,000	1	0
LiveCodeBench v6 (875 attempts)	100,000	225	149

Research status and limitations

This is an experimental research release. Swift 1.5 Ornith 9B thinks less than the base on every benchmark above, but it also scores 0.7 to 2.3 pp lower than the base on AIME, IFBench and LiveCodeBench. Only GPQA-Diamond, the benchmark we used to select this release, is higher.

We have not evaluated it on agentic, tool-use or vision benchmarks, and benchmark scores may not translate into reliable general-purpose behavior. These observations concern the specific model, runtimes and tests we used. This release shares a research result and its current limitations, rather than presenting a production-ready model.

GGUF files

File	Size	Contents
Swift-1.5-Ornith-9B-BF16.gguf	18.4 GB	language model incl. MTP head, lossless BF16
Swift-1.5-Ornith-9B-Q8_0.gguf	9.8 GB	language model incl. MTP head, Q8_0
Swift-1.5-Ornith-9B-Q6_K.gguf	7.6 GB	language model incl. MTP head, Q6_K
Swift-1.5-Ornith-9B-Q5_K_M.gguf	6.6 GB	language model incl. MTP head, Q5_K_M
Swift-1.5-Ornith-9B-Q4_K_M.gguf	5.8 GB	language model incl. MTP head, Q4_K_M
mmproj-Swift-1.5-Ornith-9B-F16.gguf	0.9 GB	vision encoder and projector, F16

The benchmark scores above were measured on the BF16 safetensors checkpoint, not on these files.

Recipe

All files were built with llama.cpp b11510 from the published BF16 safetensors: convert_hf_to_gguf.py for the BF16 file and the F16 mmproj, then llama-quantize without an importance matrix for the other tiers. Before upload, every file passed a greedy smoke test, Q8_0 with the mmproj passed a vision test, and all language-model files were checked for identical tensor names and shapes.

How to use

Use a current llama.cpp build with Ornith-1.5-9B (qwen35) support.

llama-server -hf ukisai/Swift-1.5-Ornith-9B-GGUF:Q4_K_M --jinja
# or, after downloading:
llama-server -m Swift-1.5-Ornith-9B-Q4_K_M.gguf \
  --mmproj mmproj-Swift-1.5-Ornith-9B-F16.gguf --jinja -c 32768 \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0

The model thinks before it answers. The files embed the same chat template that Transformers and vLLM use (chat_template.jinja from Ornith-1.5-9B), so --jinja gives the same prompts as the safetensors checkpoint. The model supports up to 262,144 tokens of context; set -c to what fits your memory. We evaluated with temperature=1.0, top_p=0.95, top_k=20, min_p=0.0.

For vLLM, SGLang and Transformers, use the safetensors checkpoint at ukisai/Swift-1.5-Ornith-9B.

License and access

Swift 1.5 Ornith 9B is a derivative of Ornith-1.5-9B (Ornith Team, MIT License). UkisAI's contribution, including the adapted weights, is licensed under the Apache License 2.0. See NOTICE for the change notice and attribution details.

The base model's MIT License still applies to the base model. This is an independent UkisAI release, not an official Ornith release or endorsement.

Citation

@misc{swift-1.5-ornith-9b,
  title  = {Swift 1.5 Ornith 9B},
  author = {UkisAI},
  year   = {2026},
  url    = {https://huggingface.co/ukisai/Swift-1.5-Ornith-9B}
}

Acknowledgements

We acknowledge Amazon Web Services for providing compute credits and infrastructure support for Swift's development, training, and evaluation.