Qwen 3.8 Max and 27B: Open-Weight AI Models Compared

Alibaba launched Qwen 3.8-Max, a 2.4T-parameter MoE model, and announced Qwen 3.8-27B for local use. Both will have open weights.

Compared to Kimi K3 (2.8T params), Qwen 3.8-Max offers a 1M context and significantly lower API costs. Qwen 3.8 Max charges $2.00 per million input tokens ($0.20 for cached hits) and $6.00 per million output tokens. Kimi K3 charges $2.90–$3.00 per million input tokens ($0.30 for cached hits) and $14.00–$15.00 per million output tokens. This gives Qwen a 2.4x–2.5x price advantage on output tokens, making it more cost-effective for high-throughput generation and iterative agent loops. Speed is comparable, but Qwen's efficiency gives it an edge for enterprise deployment. For self-hosting, Kimi K3 provides open model weights (594 GB MXFP4 quant), allowing organizations to eliminate API token costs entirely by running the model on private multi-node infrastructure or high-memory workstations.

In terms of performance, Kimi K3 excels at complex agentic reasoning and software engineering, leading on benchmarks like FrontierSWE (81.2% vs. 73.5%), DeepSWE 1.1 (69.0% vs. 56.6%), and Terminal-Bench 2.1 (88.3 vs. 86.6). It also holds the overall lead in academic reasoning, outscoring Qwen on GPQA Diamond (93.5% vs. 92.6%), Humanity's Last Exam (56.0% vs. 43.6%), and MLS-Bench Lite (48.3% vs. 41.0%). However, Qwen 3.8 Max leads in multimodal vision and perception, scoring higher on PerceptionBench (63.5 vs. 58.5) and MMMU-Pro (82.3% vs. 81.6%), and outperforms Kimi K3 in computer and OS automation with a verified SOTA score of 86.1% on OSWorld.

The 27B model, successor to 3.6-27B, runs quantized on a single GPU (17GB VRAM), replacing the older model for private deployments. This makes it ideal for on-premises agentic work, offering a balance of performance and privacy.

Open weights promise to democratize frontier AI. The open-weight release is expected within a week.