Qwen 3.8 and Gemma 4

Alibaba and Google

Open-weight model · runs on-premises · Published on 5 July 2026 · Updated on 2 September 2026

Qwen 3.8 and Gemma 4 are the compact models. Qwen 3.8 27B and Gemma 4, from 2 to 31 billion parameters, run on a single card, a workstation or at the edge. They suit small businesses.

Qwen3.8-27B, put online in mid-August 2026 under the Apache 2.0 licence (Hugging Face repository created on 5 August, last updated on 14 August 2026; 64 layers, 27.8 billion elements, 56 GB of weights in BF16), brings two structural changes for on-premises deployment: a native 262,144-token context (extendable to 1 million via YaRN) and a hybrid architecture, 48 Gated DeltaNet linear-attention layers and only 16 full-attention layers, which cuts the KV cache to 64 KiB per token in FP16, four times less than an equivalent classic dense model. Details and hardware sizing are covered in a dedicated article.

Key points: Compact · Qwen 3.8 27B and Gemma 4 · Single card · Edge · Small businesses.

Architecture

Qwen 3.8 27B and Gemma 4, from 2 to 31 billion parameters, are the platform's compact models. Gemma 4 comes in E2B, E4B, 12B, 26B-A4B and 31B, under the Apache 2.0 licence (google/gemma-4 repositories re-read on 2 September 2026).

Strengths

They run on a single card, a workstation or at the edge, with a reduced memory footprint.

Use cases

These models suit small businesses, embedded uses and tasks where local latency and cost matter more than raw power.

Deployment

Served by llama.cpp or vLLM on DGX Spark, Mac Studio or a single GPU card.

What about Qwen3.8-Max-Preview?

Alibaba previewed Qwen3.8-Max on 19 July 2026, a multimodal mixture-of-experts model announced at 2.4 trillion parameters. Since 8 August 2026, the base model Qwen3.8-2.4T-A95B is published on Hugging Face: 2.4 trillion parameters of which 95 billion active, a native context of 262,144 tokens, 4,892 GB of weights in BF16, under the conditional Qwen3.8-Max licence; the multimodal Max version remains served by Alibaba's API. On paper the base model can therefore be deployed on-premises, but 2.4 trillion parameters exceed every single-node machine in the catalogue, even at 4 bits. The range retained here remains Qwen3.8-27B, whose specification is published, whose weights can be downloaded and whose footprint fits a single card. See our analysis of Qwen3.8-Max.

All the models on the platform are interchangeable through a single gateway: switching models is just changing one configuration line, with no code rewrite. Local execution remains the priority.

The Qwen3-Embedding-4B variant, served in FP8 by vLLM on an RTX 5080, was measured in real conditions on 31 July 2026: 14,100 tok/s sustained and 19,075 tok/s on long text, with the full procedure in the measurement page of the measurements section.

Official sources

• Qwen/Qwen3.8-27B, official model card on Hugging Face: licence, context window and weight format
• google/gemma-4-31B-it, official model card on Hugging Face: Apache 2.0 licence
• Artificial Analysis, model comparison index