Qwen3.8 27B

PrismML released Ternary Bonsai 2 27B, a ternary-weight derivative of Qwen3.8 27B that compresses the 53.80 GB FP16 model to just 5.93 GB while retaining 98.2% of the parent's performance across 20 benchmarks. The 27.36B-parameter architecture keeps Qwen3.8 27B's hybrid attention backbone (roughly 75% linear, 25% full) and supports text plus image input at a 262K-token context length. Released under Apache 2.0, the quantized model targets local deployment on a 16 GB laptop or single RTX 5090 GPU, requiring PrismML's llama.cpp fork or MLX runtime. PrismML demonstrates it driving Cline coding agents and computer-use workloads locally, positioning the ternary variant as a practical quantization step two months after the original Bonsai 27B, which retained about 95% of Qwen3.8 27B.

Alibaba's Qwen team released Qwen3.8-27B on August 14, 2026, at 15:00 UTC as a 27.78-billion-parameter dense multimodal model accepting text, images, and video. It ships under Apache 2.0 with a native 262,144-token context window extendable to roughly 1M tokens via YaRN scaling. Qwen reports large gains over Qwen3.6-27B on agentic coding, computer use, and vision-language tasks without increasing decoder size. The official model card shows Terminal-Bench 2.1 rising from 63.4 to 73.0, DeepSWE 1.1 from 13.3 to 42.2, OSWorld-Verified from 63.9 to 84.3, and SWE-MM from 25.7 to 38.6. Kingy.ai flags that all launch scores come from Qwen, several benchmarks are in-house or modified, and no independent reproduction existed at review time. Static YaRN can also hurt shorter prompts.

Artificial Analysis profiles Qwen3.8 27B (xhigh) as an open-weight release from Alibaba dated August 2026, scoring 34 on the Intelligence Index with a 256K-token context window. The model is multimodal, accepting text, image, and video input while producing text output, and ships under Apache 2.0 with a reasoning configuration available alongside any non-reasoning variant. The benchmarking page flags Qwen3.8 27B as notably expensive relative to peer open-weight models of similar size, with $0.50 per 1M input tokens and $3.00 per 1M output tokens. Throughput is also low at 44 tokens per second, while verbosity is high at 200M tokens generated across the Intelligence Index suite, signaling strong intelligence but heavier compute and cost footprints for users.

Qwen3.8-27B is a 27.78-billion-parameter dense multimodal model released August 14, 2026 by Alibaba's Tongyi Lab under Apache 2.0. Architecture uses 64 transformer blocks with hidden dimension 5,120, 24 query heads and 4 KV heads, head dimension 256, and a 248,320-token multilingual vocabulary. Native context is 262,144 tokens, extendable to 1,048,576 tokens with YaRN. The model uses a hybrid 3:1 attention scheme with 48 Gated DeltaNet linear-attention layers and 16 gated full-attention layers, reducing complexity for 75% of layers while preserving quality on long contexts. A built-in multi-token prediction head supports speculative decoding, and a fused q_proj attention output gate operates at 12,288 by 5,120 dimensions. Vision encoder handles text, image, and video.