Releases·Model Release
UkisAI Swift-Qwen3.8-27B: -58.3% thinking tokens, 1.95x speed, xhigh accuracy kept
UkisAI post-trained Qwen3.8 27B into Swift-Qwen3.8-27B by using RL to penalize reasoning-marker tokens linked to overthinking, then fixing accuracy with on-policy distillation. The open-source model cuts thinking tokens 58.3% and speeds inference 1.95x while keeping xhigh accuracy, and community benchmarks independently confirm the gains. A thank-you post reports 100k+ downloads, #1 finetune and #9 model on HuggingFace Trending, with a Swift1.5 checkpoint and Swift Qwen3.8 Flash Next upcoming.
3485 engagement4 sourcesMon, Sep 14, 2026, 03:57 PM
Developments
Model surpassed 100k downloads, becoming #1 finetune and #9 model on HuggingFace Trending; Swift1.5 and Flash Next announced.
Thu, Sep 17, 2026, 07:30 PM
AI Overview
- UkisAI post-trained Qwen 3.8 27B to penalize overthinking-linked reasoning-marker tokens, partly post-trained on traces from BottleCap AI's ThinkingCap-Qwen3.6-27B, reporting 30-50% fewer reasoning tokens at equal quality
- UkisAI plans to release Swift Qwen3.8 Flash Next within the week, running expanded benchmarks including coding and long-horizon before publishing results.
- The team says the methodology may generalize across the Qwen family and invites community feedback on quantization and model requests.
- A Reddit user's independent Aider evals verified the claims: Swift cut completion tokens 7,301 vs 12,547, halved seconds/case (750 vs 1,481), Pass1 30.8 vs 27.1
- Swift-Qwen3.8-27B passed 100k downloads, ranking #1 finetune and #9 model on HuggingFace Trending; Swift1.5 checkpoint with training fixes lands soon
Entities
Related
Social·Community
UkisAI releases Swift 1.5 Qwen3.8 27B tuned for token efficiency; tester says IQ4_XS beats Gemini Flash on scripting task
UkisAI released Swift 1.5 Qwen3.8 27B, a Qwen-based 27B model tuned for token efficiency, on Hugging Face. A Reddit user benchmarked the IQ4_XS quant against Unsloth's Q4_K_S and Gemini Flash, reporting the 27B fixed a custom script Gemini Flash failed on. The result suggests small efficiency-tuned local models can outperform cloud models on low-reasoning coding tasks.
UkisAI released Swift 1.5 Qwen3.8 27B, a Qwen-based 27B model tuned for token efficiency, on Hugging Face. A Reddit user benchmarked the IQ4_XS quant against Unsloth's Q4_K_S and Gemini Flash, reporting the 27B fixed a custom script Gemini Flash failed on. The result suggests small efficiency-tuned local models can outperform cloud models on low-reasoning coding tasks.
85 engagement1 sourceredditSun, Sep 27, 2026, 08:46 PM
Social·Community
Swift 1.5 + HyperQwen = 37% less task completion time at 100+ tps w/ 150k context on RTX 3090
A Reddit user adapted UkisAI's Swift 1.0/1.5 finetunes of Qwen3.8 27B for syv-ai's HyperQwen speculative-decoding runtime and benchmarked them on a single RTX 3090 24GB. Swift 1.5 INT4-heads finished tasks ~37% faster than HyperQwen's W4A16 fast quant while keeping near-identical benchmark quality. Models and setup instructions are published on Hugging Face.
A Reddit user adapted UkisAI's Swift 1.0/1.5 finetunes of Qwen3.8 27B for syv-ai's HyperQwen speculative-decoding runtime and benchmarked them on a single RTX 3090 24GB. Swift 1.5 INT4-heads finished tasks ~37% faster than HyperQwen's W4A16 fast quant while keeping near-identical benchmark quality. Models and setup instructions are published on Hugging Face.
59 engagement1 sourceredditMon, Sep 28, 2026, 08:52 PM
News·Company News
UkisAI's Swift models top 2.2M downloads; lab launches Discord early access and free-compute research program
Jovan from UkisAI announced that its open-source Swift reasoning-efficient LLMs crossed 2.2 million downloads. The lab is opening a Discord with early access for the first 100 people to unreleased Swift GLM 5.3 Flash and Swift 9B models, UkisAI Code, and Swift.cpp, plus a research support program offering free compute, APIs, and datasets.
Jovan from UkisAI announced that its open-source Swift reasoning-efficient LLMs crossed 2.2 million downloads. The lab is opening a Discord with early access for the first 100 people to unreleased Swift GLM 5.3 Flash and Swift 9B models, UkisAI Code, and Swift.cpp, plus a research support program offering free compute, APIs, and datasets.
253 engagement1 sourceredditThu, Oct 8, 2026, 03:50 PM
Social·Community
peculiar-ragdoll's Dirk-Qwen 3.8-27B beats UkisAI Swift-1.5 Qwen3.8-27B on personal domain-specific eval
A Reddit user ran their own domain-specific eval suite (~200+ questions spanning coding, data analytics, local RAG, and voice assistant tasks) comparing two 27B Qwen3.8-based local models on an M1 Max. peculiar-ragdoll's Dirk-Qwen3.8-27B substantially outperformed UkisAI's Swift-1.5, producing sharper decisions, more correct answers, and lower token usage. The result matters to local-model practitioners choosing community fine-tunes for daily workflows.
A Reddit user ran their own domain-specific eval suite (~200+ questions spanning coding, data analytics, local RAG, and voice assistant tasks) comparing two 27B Qwen3.8-based local models on an M1 Max. peculiar-ragdoll's Dirk-Qwen3.8-27B substantially outperformed UkisAI's Swift-1.5, producing sharper decisions, more correct answers, and lower token usage. The result matters to local-model practitioners choosing community fine-tunes for daily workflows.
34 engagement1 sourceredditWed, Sep 30, 2026, 10:14 PM
Developments·Benchmark
ThinkingCap 3.8-27B vs. Swift 3.8-27B vs. Qwen 3.8-27B Benchmarks
Multiple LocalLLaMA users benchmarked two fine-tunes of Qwen3.8-27B — BottlecapAI's ThinkingCap and ukisai's Swift — that both claim ~40% reasoning-token reduction. Independent Aider-eval and 7900 XTX tests largely confirm the claims: median completion tokens drop ~5k with minimal quality loss. Trade-offs differ: Swift has the fastest prefill, ThinkingCap the slowest but the best-formed diffs, and both show weaknesses in edge cases and per-language results.
Multiple LocalLLaMA users benchmarked two fine-tunes of Qwen3.8-27B — BottlecapAI's ThinkingCap and ukisai's Swift — that both claim ~40% reasoning-token reduction. Independent Aider-eval and 7900 XTX tests largely confirm the claims: median completion tokens drop ~5k with minimal quality loss. Trade-offs differ: Swift has the fastest prefill, ThinkingCap the slowest but the best-formed diffs, and both show weaknesses in edge cases and per-language results.
Updated Thu, Oct 1, 2026, 02:52 PM — A follow-up eval added unsloth Q4_K_XL and Peculiar-Ragdoll checkpoints to the Swift1.5/ThinkingCap comparison over 69 q
110218 engagement3 sourcesredditThu, Sep 24, 2026, 04:11 PM