Qwen3.8 27B

On September 18, 2026, PrismML released Ternary Bonsai 2 27B, a derivative ternary-weight quantization of Qwen3.8 27B that preserves the parent's architecture. MarkTechPost reports the model occupies 5.93 GB versus 53.80 GB in FP16 and retains 98.2% of Qwen3.8 27B's average across 20 benchmarks, while accepting text an Ternary weights cover embeddings, attention projections, MLP projections, and the LM head, with only 26.2M parameters (0.0976%) kept in higher precision along the recurrent state path and normalization weights; the vision tower ships separately as a 0.63 GB GGUF file. The Apache 2.0 weights are positioned for deploymen

Kingy.ai's launch-day review confirms that Alibaba's Qwen team released Qwen3.8-27B at 15:00 UTC on August 14, 2026, as a 27.78-billion-parameter multimodal checkpoint under Apache 2.0 with a native 262,144-token context and text, image, and video inputs. The page notes Qwen-reported gains over Qwen3.6-27B in agentic c The review adds important caveats: every launch score is sourced from Qwen, several benchmarks are in-house corrected or modified, and the SWE-bench Pro comparison imports Anthropic's Opus result rather than rerunning it. Kingy.ai flags that the most aggressive local-hardware claims ignore KV cache, that the 1M-token w

On August 14, 2026, Alibaba's Tongyi Lab released Qwen3.8-27B as a 27.78-billion-parameter dense multimodal language model under Apache 2.0, sized for local deployment (minimum 24GB VRAM). The Local AI Zone technical analysis describes a hybrid attention architecture built on the Qwen3.5 foundation: 64 Transformer bloc The same analysis reports a 3:1 layered attention ratio combining 48 Gated DeltaNet linear-attention layers with 16 Gated Full Attention layers, plus an output gate fused into a 12,288×5,120 q-projection. Local AI Zone summarizes Qwen-reported results showing Qwen3.8-27B outperforming Meta's Muse Glimmer (30B) across a

Alibaba's Qwen team released Qwen3.8-27B as an open-weight dense 27-billion-parameter vision-language model under Apache 2.0, featuring a native 262,144-token context window with an advertised extension path to one million tokens. According to NxCode's analysis, the published configuration is architecturally identical The model is available through multiple serving targets: official FP8 weights for vLLM, SGLang, Transformers, and TokenSpeed, plus Ollama packages listing 18GB for both standard and MLX variants aimed at Apple Silicon workstations. Qwen's vendor-run tables report agent-evaluation gains over Qwen3.6-27B on Terminal Benc