license: other
license_name: swift-open-license-1.0
license_link: LICENSE
library_name: gguf
pipeline_tag: text-generation
base_model: ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF
base_model_relation: finetune
tags:

• gguf
• llama.cpp
• qwen3_8
• gsq
• rco
• reasoning
• swift
• gsq-rco
• mtp
• uncensored
• decensored
• chips

Swift 1.5 Qwen3.8-27B · CHiPS-Uncensored · GSQ-RCO

Modified versions of UkisAI's IQ2_XS-mtp and IQ3_S-mtp files from Swift 1.5 Qwen3.8-27B GSQ-RCO, with refusal behavior reduced.

Not an official UkisAI release. UkisAI, Alibaba Cloud and ISTA-DASLab did not make it and do not endorse it. "Swift" is used only to describe where the model comes from.

All numbers are for the IQ2_XS-mtp file. The IQ3_S-mtp file has the same edit but was not evaluated.

• Refusal phrases in the answers to 104 harmful requests drop from 93.3% to 2.9%.
• Loops, empty answers, long-task behavior, tool calling and perplexity stay at the level of the original file.
• Accuracy on 250 GSM8K and 300 MMLU-Pro questions also stays at that level: 94.0% against 94.8%, and 73.7% for both.
• On harmful requests answers are much longer than the original's short refusals, and 13.5% of them reach the 1,536-token limit used in the test.
• On hard questions the reasoning can run a little longer: 32 of the 300 MMLU-Pro questions reach the 4,096-token limit, against 25 for the original file.
• The edit was made with CHiPS (see below). The recipe is not published.

What changed

The rewritten tensors were requantized to their original types with the original repository's Swift V1MIX importance matrix (imatrix-swift15-v1mix.gguf).

About CHiPS

CHiPS (Causal Head-isolated Projection Steering) changes one behavior by editing only a few attention heads, chosen by their causal effect on that behavior. The rest of the model is left untouched. The recipe is not published.

Files

Each file has the same quantization as the original build of its tier, including the MTP head. LICENSE, LICENSE-APACHE-2.0 and NOTICE hold the license terms and attribution (see below).

Only these two tiers are provided. No vision projector is included, and vision was not tested.

Evaluation

Every number below was measured on the released IQ2_XS-mtp file and on the original IQ2_XS-mtp file (verified against the original repository's SHA-256), one after the other, with the same settings and seeds. Each set was run once. The IQ3_S-mtp file was not evaluated: the edit was tuned on IQ2_XS-mtp, and requantizing can change how much of a small edit survives, so its behavior may differ.

Requests

104 harmful and 100 harmless single-turn requests in English and Japanese, at most 1,536 new tokens. The request sets are not published.

An empty answer means the reasoning did not finish within the limit. On harmful requests none of the answers is empty, so all 14 that reached the limit had started answering. If you want complete long answers, raise max_tokens.

Long tasks and tool calling

24 long-form tasks (at most 3,072 new tokens) and 40 tool-calling tasks (at most 1,536 new tokens per turn; 30 expect a call and 10 do not).

Differences of one or two tasks are within what sampling alone produces, so read this table as "the same level", not as an improvement.

Capability

250 GSM8K and 300 MMLU-Pro questions, drawn from the test splits with a fixed seed, at most 4,096 new tokens. Both files answer the same questions. A reply without a final answer counts as wrong.

The accuracy differences are within what sampling produces (exact two-sided McNemar test: p = 0.69 for GSM8K, p = 1.00 for MMLU-Pro). On MMLU-Pro every reply without an answer reached the token limit, and a higher limit would let more of them finish. Of the questions that reached the limit on only one file, 11 were on this model and 4 on the original (p = 0.12), so this could be sampling too, but it is the one measure here that leans against this model.

These accuracies depend on the settings below and on the sampled questions, so do not compare them with figures from other reports.

Perplexity

Teacher-forced perplexity on small in-house reference texts, in 512-token chunks. These are not comparable with the KLD figures in the original card below.

All differences are within 0.5%.

Evaluation setup

• llama.cpp (llama-server), thinking enabled (reasoning_effort: medium), temperature 0.6, top_p 0.95, top_k 20, presence penalty 0, fixed seeds, q8_0 KV cache, 4 parallel slots.
• Refusals are counted by matching common refusal phrases (English and Japanese) in the final answer.
• A short-period loop is heavy repetition of short character sequences at the end of the output. A long-period loop is the second half repeating chunks of the first half.
• Tool results are fixed placeholder values.
• For the benchmarks, the answer is taken from the last Answer: ... line or \boxed{} in the final answer.

Usage

hf download Kujira/Swift-1.5-Qwen3.8-27B-CHiPS-Uncensored-GSQ-RCO-MTP-GGUF Swift-1.5-Qwen3.8-27B-CHiPS-Uncensored-GSQ-RCO-IQ2_XS-mtp.gguf --local-dir .

llama-server -m Swift-1.5-Qwen3.8-27B-CHiPS-Uncensored-GSQ-RCO-IQ2_XS-mtp.gguf -ngl 99 -fa on --jinja

For IQ3_S, replace IQ2_XS with IQ3_S in both commands. The files are standard GGUFs and need no adapter. Use a llama.cpp build that supports Qwen3.8. The sampling settings in the evaluation setup are a good starting point.

Limitations

• Refusals were counted by phrase matching and the answers were not reviewed by people. An answer that avoids the request without using a refusal phrase is not counted as a refusal, and a phrase can also appear in an answer that does comply.
• Each set was run once and the sets are small: one request is about 1%, and one long task about 4%.
• Only single-turn requests without a system prompt were tested. Multi-turn conversations, system prompts, and rephrased or adversarial requests were not.
• Capability was measured with GSM8K and MMLU-Pro only, on one sample of each. Other benchmarks were not run, and the KLD results in the original card were not re-measured.
• The IQ3_S-mtp file was not evaluated at all.
• The MTP head is unchanged, but speculative decoding was not tested.

Intended use

This model follows many requests that the original model refuses, and it can produce harmful, offensive or inaccurate content. It is meant for research and for users who understand this. You are responsible for how you use it, including compliance with applicable laws and the license below. Add your own safeguards before using it in anything that serves other people.

License

Same license as the original release: the Swift Open License v1.0, including its commercial-use limitation (Section 5). The Qwen components retain the Apache License 2.0. See NOTICE for attribution and for the changes made in this repository.

Copyright 2026 UkisAI. Swift Contribution licensed under the Swift Open License v1.0 (https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b/blob/main/LICENSE).
Derivative of Qwen3.8-27B, Copyright 2026 Alibaba Cloud, Apache License 2.0.

Credits

• Swift 1.5 and this GGUF release: UkisAI.
• Base model: Qwen3.8-27B by the Qwen team.
• Quantization methods: GSQ and RCO by ISTA-DASLab (see the original card below).

Original model card

The following is the model card of ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF, reproduced for reference without its banner. The files, results and usage below refer to the original repository, not to this model.

Swift 1.5 Qwen3.8-27B · GSQ-RCO

Compact, mixed-precision GGUF quantizations of Swift 1.5 Qwen3.8-27B, with Swift-specific refinement using the per-tensor allocations from ISTA-DASLab's GSQ-RCO release.

Swift 1.5 builds on Swift 1.0 through post-training focused on long-horizon, agentic and coding tasks, improving overall performance while using fewer thinking tokens. See the original model card for the model's training approach and benchmark results. Those model-level benchmarks are separate from the quantization measurements below.

Swift 1.5 uses 58.5% fewer thinking tokens while scoring 0.35% higher than the base, for a 9.18× speed-up on several tasks.

Available quantizations

Each tier is a single GGUF file. Sizes are decimal GB; runtime memory also includes the context cache and compute buffers. Tier names denote mixed-precision allocation profiles, rather than a uniform type for every tensor.

The optional -mtp files retain the matching refined model tensors and add the MTP head. They require a runtime with support for this model's MTP implementation. The KLD results here were measured on the standard files; MTP decoding speed and quality have not been separately evaluated.

Exact file identities are recorded in release-manifest.json and SHA256SUMS.

Evaluation

KLD measures the divergence of the quantized model's next-token distribution from Swift 1.5 BF16; lower is better. Development measurements use wiki.test.raw, 100 chunks and a 512-token context. Development data informed refinement and is not independent validation.

The held-out results below use C4 prose, CodeParrot code, GSM8K math text and multilingual mC4 text. Prose/code/math use 100 chunks each; German/French/Spanish/Chinese use 25 chunks each, all at context 512. These are distributional KLD measurements, not task accuracy or math benchmark scores.

All four refined files improve KLD over their matched Swift starting quantizations on all seven reporting domains. Results are not uniformly better than the ISTA comparison quants: math-text KLD is 4.9–7.7% higher, and IQ3_S is higher on five of seven domains. ISTA comparisons measure each quant against its own corresponding BF16 model; they are not direct Swift-versus-Qwen task rankings or proof of equivalent capability.

The full result table includes the matched Swift starting quant, ISTA comparison and reported error estimates. Evaluation metadata binds results to the exact released identities. A lexical overlap screen was applied against calibration/development text; it does not establish semantic deduplication or prove absence of overfitting. These 512-token tests do not establish quality at 32K or longer contexts.

Usage

Use a llama.cpp build that supports Qwen3.8. Authenticate with an account granted access while this repository is private.

hf download ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF Swift-1.5-Qwen3.8-27B-GSQ-RCO-IQ3_XXS.gguf --local-dir .

llama-server \
  -m Swift-1.5-Qwen3.8-27B-GSQ-RCO-IQ3_XXS.gguf \
  --jinja -fa on -ngl 99 -c 262144 \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 \
  --presence-penalty 0.0 --repeat-penalty 1.0 --port 8000

Set the context size to fit available memory. The example context setting is not a claim that these quants were evaluated at that length.

This release provides the language-model GGUFs. A Swift 27B vision projector has not been verified for this release, so no projector or validated vision example is included.

Quantization procedure

1. Reuse the published ISTA GSQ-RCO per-tensor allocation for each matching Qwen3.8-27B tier.
2. Quantize Swift 1.5 weights with the Swift V1MIX importance matrix and the selected allocation.
3. Apply Swift-specific GSQ refinement. IQ2_XS and IQ3_XXS use the preserved fixed-objective variants; IQ2_S and IQ3_S use their preserved V1MIX variants.
4. Freeze exact file identities, run the reporting evaluations, and preserve the matching MTP packages.

This release reuses ISTA's allocation search results; it does not claim a new RCO search on Swift. The archived recipe records document construction settings and historical candidates; the release manifest identifies the selected files. The Swift importance matrix is included as imatrix-swift15-v1mix.gguf.

Per-tensor assignments for all eight files are included in tensor-allocation. Each dump records the model SHA256, tensor count and type histogram. MTP dumps contain the 851 model tensors plus 15 head tensors.

Methods and acknowledgements

GSQ and RCO were developed by the Deep Algorithms and Systems Lab at the Institute of Science and Technology Austria. This Swift adaptation is by UkisAI.

• GSQ: paper · code
• RCO: paper · code
• GGUF runtime and conversion: llama.cpp

We acknowledge the Qwen team for the original model, ISTA-DASLab for the quantization methods and published allocations, and NVIDIA Innovation Lab, AWS and Google Cloud for supporting Swift development.

License and access

These adapted weights are distributed under the Swift Open License v1.0. The original Qwen components retain their Apache 2.0 license and NOTICE. See the Swift license for applicable terms and enterprise licensing information.