Qwen3.8 27B (NovitaAI)

Released on August 14, 2026, Qwen3.8 27B is a 27B Apache 2.0 multimodal model that passed one million downloads in two days and can fit a 17GB quantized build on suitable consumer hardware, according to this research answer. Its dense 27B design keeps all model parameters resident, while 48 of its 64 layers use linear The model is described as open-weight, licensed under Apache 2.0, dense rather than MoE, and natively multimodal for text, images, and video, with a 262,144-token native context window and longer contexts possible through YaRN. The article frames Qwen3.8-27B as making a frontier-adjacent capability tier practical on ha

This community deep-dive describes Qwen3.8-27B as the compact, deployment-friendly member of the new Qwen3.8 generation, built on the architecture Qwen introduced in Qwen3.5. It is a causal language model with a vision encoder natively trained to reason over text, images, and hour-long videos. Key features include codi The architecture breakdown specifies 27B parameters (28B with padding), hidden dimension 5,120, 64 layers with a 16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN)) pattern, 48 Gated DeltaNet heads (V) / 16 (QK) with head dim 128, 24 Gated Attention heads (Q) / 4 (KV) with head dim 256 and RoPE dim 64, FFN

Qwen3.8-27B is a dense 27B open-weights vision-language model released August 14, 2026 under Apache-2.0, scoring 52 on Artificial Analysis' Intelligence Index—up 14 points over Qwen3.6-27B on an identical architecture—according to this guide. It fits in an 18 GB Ollama download and is positioned as a strong local codin Specs include a 17 GB Q4-quantized file, native image and video input including hour-scale video, 262,144-token native context extensible to 1M (the 1M figure being a Qwen Cloud hosted feature marked 'coming soon'), and a reasoning effort control with xhigh as default. The architecture is identical to Qwen3.6-27B—same

DataNorth reports that Alibaba released Qwen3.8-27B on 14 August 2026 as an open-weights dense model accepting text, images, and video, shipped under the Apache 2.0 license. The model carries a 262,144-token native context window that Alibaba extends to 1,000,000 tokens, and is the smaller companion to Qwen3.8-Max (who DataNorth emphasizes that Qwen3.8-27B is a native vision-language model rather than a text model with a bolted-on adapter, with the configuration file setting the language-model-only flag to false and shipping runnable image and video examples in the model card. Alibaba also ships an FP8 variant, Qwen3.8-27B-FP8, using

Qwen3.8-27B is a 27-billion-parameter vision-language model released by Alibaba's Qwen team on Hugging Face around August 5-14, 2026, under Apache-2.0 licensing. Most attention layers use a Gated DeltaNet linear-attention variant, enabling a native 262,144-token context window without standard memory blow-up, while accepting text, images, and hour-long video input. The model topped Hacker News within a day of release and became the most-liked model on Hugging Face's trending list that week. Qwen's own benchmarks show it beating Claude Opus 4.6 Max on computer-use and browser-use tasks, though it trails top closed models on raw reasoning benchmarks such as HLE and GPQA Diamond.

This third-party technical analysis confirms Qwen3.8-27B was released on August 14, 2026, at 15:00 UTC by the Qwen Team at Alibaba Cloud (Tongyi Lab) under Apache 2.0 license. The model is described as a 27.8-billion-parameter dense multimodal language model with a hidden dimension of 5,120, 64 transformer blocks, 24 q The analysis details a hybrid attention mechanism with a 3:1 ratio: 48 layers (75%) use Gated DeltaNet linear attention while 16 layers (25%) use gated full attention, with an attention output gate fusing q_proj at 12,288 × 5,120 dimensions. It claims the model outperforms Meta's Muse Glimmer (30B) across all 8 direct

Alibaba's Qwen team released Qwen3.8-27B at 15:00 UTC on August 14, 2026, according to this independent technical review. The model checkpoint contains 27.78 billion parameters, accepts text, images and video, ships under Apache 2.0, and has a native 262,144-token context window. The review notes large reported improve The review also flags caveats around local-hardware claims: one million tokens require YaRN scaling rather than the native context, and Qwen warns that static YaRN can hurt shorter prompts. Kingy.ai inspected the model card, configuration, license, repository metadata, official BF16 and FP8 artifacts, and third-party G

Artificial Analysis published independent benchmarks for Qwen3.8 27B on August 17, 2026, assigning an Intelligence Index score of 52, matching GPT-5.6 Luna. The model's Agentic Index score of 51 surpassed both GPT-5.6 Terra and Claude Opus 4.8, demonstrating agent capabilities well above its 27B parameter scale. The open-source model runs locally on high-end consumer hardware via tools such as Ollama and LM Studio. Despite its compact size relative to frontier models, Qwen3.8 27B delivers competitive performance, positioning it as a notably efficient option for local agentic and coding workflows compared to larger closed alternatives.