Qwen 3.8 27B
A 27-billion-parameter open-source model in Alibaba's Qwen series, built on the Qwen 3.5 architecture. It natively supports vision-language multimodal input and employs a hybrid attention design combining gated DeltaNet and gated attention mechanisms, with optimization targeting local AI agent deployment scenarios.
Core Facts
Timeline (last 90 days)
The user's previous main model was Qwen 3.8 27B, using Q6 quantization and the kv8 configuration.
When asked in testing about the best barbecue in Austin, Qwen 3.8 27B listed 5 places with JSON and 6 with Markdown, with the added Good BBQ Company being one of the two highest-rated.
On a MacBook Pro M5 Pro with 48 GB of memory and a 32K context window, the average time per search question in Markdown (restrictor) mode: Qwen 3.8 27B about 42 seconds, Meta Muse Glimmer 30B about 62 seconds, Prism Bonsai 27B about 24 seconds, and Gemma 4 12B QAT about 16 seconds (about 43 seconds after switching to Markdown Raw).
The Qwen Image model file takes up as much as 25GB, requiring sufficient disk space to be reserved.
Jordan Ilchif tested Qwen 3.8 27B on a 32GB M4 MacBook Air, with Ollama's NVFP4 build running at 17.2 tokens/s, pure MLX at 6.6 tokens/s, and LLAMA.cpp at 4.4 tokens/s, but the formats were different.
Vercel AI Gateway switches to Ling 3.1 Flash, Tempr Gateway changes to MiMo V2.6 Flash, CrossModel and Ofox move to Claude Sonnet 5.5, Pioneer adopts GLiNER 2.5 Decide, and Scaleway uses Qwen 3.8 27B.
A developer connected an iPhone 17 Pro Max to a 24GB-memory M4 Pro MacBook via a 10 Gb/s USB-C cable, using the phone as a second GPU to share the computational load for a locally running Qwen 3.8 27B (IQ4_XS quantization) model.
The open-source model Qwen 3.8 27B (17GB) running on Willison's laptop was the first model runnable on a laptop that felt genuinely competitive with frontier models.
It is recommended to use Low or Medium reasoning levels for mechanical steps and reserve X-High for parts that truly require deep thinking, in order to save time.
Qwen 3.8 27B adopts a hybrid architecture in which, across 64 layers, the majority use Gated DeltaNet layers with Gated Full Attention layers interspersed at regular intervals.
12 more timeline events
All Facts (17)
A 27B-parameter model can run in quantized form on a consumer-grade GPU with 24GB of VRAM (such as the NVIDIA RTX 4090), without requiring data-center-level hardware.
65%UnverifiedQwen 3.8 27B adopts a hybrid architecture in which, across 64 layers, the majority use Gated DeltaNet layers with Gated Full Attention layers interspersed at regular intervals.
60%UnverifiedQwen 3.8 27B adopts an open-weight model with permissive licenses such as Apache 2.0, allowing commercial use and derivative development.
60%UnverifiedThe user's previous main model was Qwen 3.8 27B, using Q6 quantization and the kv8 configuration.
50%UnverifiedOn a MacBook Pro M5 Pro with 48 GB of memory and a 32K context window, the average time per search question in Markdown (restrictor) mode: Qwen 3.8 27B about 42 seconds, Meta Muse Glimmer 30B about 62 seconds, Prism Bonsai 27B about 24 seconds, and Gemma 4 12B QAT about 16 seconds (about 43 seconds after switching to Markdown Raw).
50%UnverifiedThe Qwen Image model file takes up as much as 25GB, requiring sufficient disk space to be reserved.
50%UnverifiedJordan Ilchif tested Qwen 3.8 27B on a 32GB M4 MacBook Air, with Ollama's NVFP4 build running at 17.2 tokens/s, pure MLX at 6.6 tokens/s, and LLAMA.cpp at 4.4 tokens/s, but the formats were different.
50%UnverifiedVercel AI Gateway switches to Ling 3.1 Flash, Tempr Gateway changes to MiMo V2.6 Flash, CrossModel and Ofox move to Claude Sonnet 5.5, Pioneer adopts GLiNER 2.5 Decide, and Scaleway uses Qwen 3.8 27B.
50%UnverifiedA developer connected an iPhone 17 Pro Max to a 24GB-memory M4 Pro MacBook via a 10 Gb/s USB-C cable, using the phone as a second GPU to share the computational load for a locally running Qwen 3.8 27B (IQ4_XS quantization) model.
50%UnverifiedThe open-source model Qwen 3.8 27B (17GB) running on Willison's laptop was the first model runnable on a laptop that felt genuinely competitive with frontier models.
50%UnverifiedIt is recommended to use Low or Medium reasoning levels for mechanical steps and reserve X-High for parts that truly require deep thinking, in order to save time.
50%UnverifiedAlibaba officially claims that the overall performance of Qwen 3.8 27B has surpassed Qwen 3.7 Plus, with particularly outstanding performance in real-world programming, agent tasks, and office workflows.
50%UnverifiedIn a Three.js game generation test, Qwen 3.8 27B produced a complete game featuring different weapons, zombies attacking in waves, unlockable areas, a mystery box purchase mechanic, and instant-death power-up items.
50%UnverifiedQwen 3.8 27B surpasses GPT-5.6 Terra, DeepSeek v4 Pro, GPT-5.6 Luna, and Gemini 3.7 Flash on the Agentic index.
50%UnverifiedArtificial Analysis gives the Qwen 3.8 27-billion-parameter model a score of 51 on the Agentic index.
50%UnverifiedWhen asked in testing about the best barbecue in Austin, Qwen 3.8 27B listed 5 places with JSON and 6 with Markdown, with the added Good BBQ Company being one of the two highest-rated.
50%UnverifiedAlibaba's Qwen team released the Qwen 3.8 27B model in open-weights form.
60%