Qwen3.8
Alibaba's current open-weight family, led for on-premises use by Qwen3.8-27B: a dense, Apache 2.0, native vision-language model small enough to run on a single Faraday node.
Model Story
The Qwen series began in 2023 as Alibaba's answer to Western frontier models and has become one of the most prolific open-weight families in AI. Qwen3.8-27B, released on August 14, 2026, is the member of the Qwen3.8 family that matters most for on-premises deployment: a dense 27 billion parameter model, licensed under Apache 2.0, with native vision-language support for both images and video and a 262K-token context window.
At the top of the family sits Qwen3.8 Max, a flagship of roughly 2.4 trillion total parameters with about 95 billion active. It is a different proposition entirely: it ships under a custom, non-Apache license and is sized for datacentre deployment, not desktop hardware. For businesses that want to own their AI stack, the 27B model is the practical choice.
Alibaba announced Qwen 4 on September 22, 2026 at its Apsara conference, with four tiers including a 27B open model. Qwen 4 has not been released and Alibaba has not published its specifications, so Qwen3.8-27B remains the current Qwen model on Faraday hardware until new weights actually ship.
For Faraday Machines customers, Qwen3.8-27B is the fast, lightweight option: everyday coding help, multilingual drafting and translation, and image or video understanding, all on a single Mac Studio.
Key Specifications
| Developer | Alibaba Cloud (Qwen team) |
|---|---|
| Release Date | August 14, 2026 (Qwen3.8-27B) |
| Architecture | Dense transformer, native vision-language |
| Parameters | 27 billion (dense, all active) |
| Context Window | 262K tokens |
| License | Apache 2.0 (Qwen3.8-27B); Qwen3.8 Max uses a custom license |
| Multimodal | Text, image, and video input |
| Size at 4-bit | Roughly 17 GB |
| Faraday Tier | Starter (1× Mac Studio M5 Max, 128GB) |
| Family Flagship | Qwen3.8 Max: ~2.4T total / ~95B active, datacentre-only |
Pricing
The Apache 2.0 license carries no licensing fee, no usage caps, and no vendor lock-in. Running Qwen3.8-27B on a Faraday Starter tier eliminates per-token costs entirely; you pay only for hardware amortization and electricity.
Benchmarks
Qwen reports 61.7% on SWE-bench Pro using its own harness, so the figure is not directly comparable with standardized runs. For context, the top score on Scale's standardized public SWE-bench Pro set is around 61.5%. Artificial Analysis scored Qwen3.8-27B at 52 on its Intelligence Index v4.1.1. Artificial Analysis rescaled the index in September 2026 (v4.3), and every model's score dropped, so this v4.1.1 score should not be compared with v4.3 scores on other profiles.
On-Premises Deployment
At 4-bit quantization Qwen3.8-27B needs roughly 17 GB of memory, which fits comfortably on a single Mac Studio M5 Max with 128GB of unified memory. That makes it the natural entry point on Faraday's Starter tier, with plenty of memory left over for long documents and other workloads.
Because it is a dense model, all 27 billion parameters are active for every token; its strength is a small footprint rather than raw scale. For heavier work, Faraday pairs it with larger open models on multi-node tiers: GLM-5.3-Flash and MiMo-V2.6-Flash on Growth, and MiniMax M3 and DeepSeek V4.1 Flash on Scale.
Qwen3.8 Max cannot be self-hosted on Faraday tiers: its size puts it in datacentre territory and its custom license differs from the Apache 2.0 terms of the 27B model. When Qwen 4 ships with published weights and specifications, its 27B open model will be the one to watch for the Starter tier.
Frequently Asked Questions
Can Qwen3.8-27B run on a single Mac Studio?
Yes. At 4-bit quantization Qwen3.8-27B needs roughly 17 GB of memory, which fits comfortably on one Mac Studio M5 Max with 128GB of unified memory. That makes it the entry model on Faraday Machines' Starter tier, built for teams of 2–9 people, with plenty of memory left over for long documents and other workloads.
Is Qwen3.8 open source?
Qwen3.8-27B is open source under the Apache 2.0 license, which carries no licensing fee, no usage caps and permits commercial use. The family flagship, Qwen3.8 Max, is different: at roughly 2.4 trillion total parameters with about 95 billion active, it uses a custom, non-Apache license and is sized for datacentre deployment, so it cannot be self-hosted on Faraday Machines tiers.
How much does it cost to run Qwen3.8-27B on-premises?
Running Qwen3.8-27B on Faraday Machines' Starter tier costs $9,999 US one-time for one Mac Studio M5 Max with 128GB of memory, including 12 months of updates, maintenance and support. The Apache 2.0 license adds no fee, and there are no per-token charges, so ongoing costs come down to hardware amortization and electricity.
Is Qwen 4 available yet?
No. Alibaba announced Qwen 4 on September 22, 2026 at its Apsara conference, with four tiers including a 27B open model, but it has not been released and Alibaba has not published its specifications. Until new weights ship, Qwen3.8-27B remains the current Qwen model on Faraday Machines hardware.
Explore Other Models
See how every model compares on price, benchmarks and hardware fit in the AI Model Comparison.