Two efficiency-focused Qwen3.8-27B derivatives trend on Hugging Face: UkisAI Swift-Qwen3.8-27B and Prism ML Ternary-Bonsai-2-27B

UkisAI's Swift-Qwen3.8-27B GGUF (trending #30 on 2026-09-11) cuts median thinking tokens by 58.3% with under 1% performance loss and a 1.95x speedup, while Prism ML's Ternary-Bonsai-2-27B (trending #29 on 2026-09-16) uses ternary {-1, 0, +1} weights at 1.72…

In mid-September 2026, two independent derivatives of Qwen3.8-27B focused on inference efficiency charted on Hugging Face. On 2026-09-11, UkisAI released Swift-Qwen3.8-27B as GGUF (trending #30), a reasoning-efficient derivative using a Swift adapter that reduces median thinking tokens by up to 58.3% (on GPQA-Diamond vs the Qwen3.8-27B base, with only a 0.10-point score drop) while keeping performance losses under 1% and delivering a 1.95x speed-up on several tasks. Its reported benchmarks include GPQA-Diamond 88.28%, MMLU-Pro 84.95%, C-Eval 90.62%, AIME 2026 94.00% and Terminal-Bench 2.1 65.84%; BF16 weights and enterprise licensing are also offered. On 2026-09-16, Prism ML published Ternary-Bonsai-2-27B (trending #29), an Apache 2.0 open-weights model derived from Qwen3.8-27B with 27.36B parameters, hybrid attention (~75% linear) enabling 262K-token context. Its weights are end-to-end ternary {-1, 0, +1} with g128 FP16 scales at 1.72 bits/weight, shrinking the model from ~54 GB FP16 to 5.95 GB (PTQ1_0) or 7.21 GB (PQ2_0) while reportedly retaining 98.2% of FP16 quality (84.78 average across 14 thinking-mode benchmarks). It ships as GGUF for llama.cpp (CUDA, Metal, CPU) with custom ternary kernels, plus an MLX 2-bit build for Apple Silicon and an optional Q8_0 vision tower pack, with a whitepaper, demo repo, and forked llama.cpp/MLX runtimes. The two releases are complementary rather than conflicting: Swift targets thinking-token overhead, while Ternary-Bonsai-2 targets weight-compression footprint.

• UkisAI released Swift-Qwen3.8-27B as GGUF on Hugging Face, trending at #30 as of 2026-09-11.
• Swift-Qwen3.8-27B uses a Swift adapter to cut median thinking tokens by 58.3% on GPQA-Diamond versus the Qwen3.8-27B base, with only a 0.10-point score drop.
• Swift-Qwen3.8-27B delivers a 1.95x speed-up on several tasks while keeping performance losses under 1%.
• Swift-Qwen3.8-27B reported benchmarks: GPQA-Diamond 88.28%, MMLU-Pro 84.95%, C-Eval 90.62%, AIME 2026 94.00%, Terminal-Bench 2.1 65.84%.
• Swift-Qwen3.8-27B is also available with BF16 weights and enterprise licensing.
• Prism ML published Ternary-Bonsai-2-27B, trending at #29 as of 2026-09-16, an Apache 2.0 open-weights model derived from Qwen3.8-27B.
• Ternary-Bonsai-2-27B has 27.36B parameters, hybrid attention (~75% linear) enabling 262K-token on-device context.
• Ternary-Bonsai-2-27B uses end-to-end ternary {-1, 0, +1} weights with g128 FP16 scales at 1.72 bits/weight, shrinking ~54 GB FP16 to 5.95 GB (PTQ1_0) or 7.21 GB (PQ2_0).