Timeline
- 8.5
Qwen3.8 27B Scores 52 on Artificial Analysis, Topping Larger Models
Qwen3.8-27B, a 27B-parameter open-weight model from Alibaba's Qwen team, scored 52 on the Artificial Analysis benchmark, outperforming models up to 150B+ parameters and matching DeepSeek V4 Flash 0731. It also rivals or beats Claude Opus 4.6, a frontier model released just six months earlier. This is a major open-source AI milestone, showing a small dense model can match or exceed much larger frontier models. It could democratize access to high-end AI capability, enabling developers to run frontier-level performance on consumer hardware. The model is a native multimodal vision-language model with Apache 2.0 open weights, excelling at coding, agentic workflows, and office automation. Community benchmarks show it scores 52, the same as DeepSeek V4 Flash 0731, and beats Opus 4.6 on Artificial Analysis; some users note Opus still has better world knowledge.
- 8.0
Qwen3.8-27B: 27B Open-Weight Model Outperforms Larger Models on Coding
Alibaba's Qwen team released Qwen3.8-27B, a 27-billion-parameter open-weight model that reportedly beats larger models such as Claude Opus on coding benchmarks while remaining small enough to run on a laptop. The Hugging Face release is the FP8-quantized version, and community GGUF quantizations are already available. This release narrows the gap between frontier models and locally runnable models, enabling developers to use advanced reasoning and coding capabilities on consumer hardware without expensive API fees. It also signals intensifying competition in the open-weight segment, which could pressure closed-model pricing. The Hugging Face release is the FP8-quantized version (Qwen3.8-27B-FP8), and community quantizations such as Unsloth's GGUF files are available for local inference via llama.cpp. In community tests, it scored 42.2 on DeepSWE, beating Claude Opus 4.7 Max's 40 when used with Claude Code.
- 8.0
Qwen Releases Qwen3.8-2.4T: 2.4T-Parameter MoE Model
Qwen released Qwen3.8-2.4T-A95B, a mixture-of-experts model with 2.4 trillion total parameters and 95 billion active parameters. The model card claims near-frontier performance comparable to Claude Opus and the Fable model family, with BF16 and FP8 weights published on Hugging Face. As a major open-weight release from a leading AI lab, Qwen3.8-2.4T pushes open models closer to proprietary frontier systems while keeping inference cost low through MoE's sparse activation. This could accelerate deployment of high-end open models in both research and production environments. The model is only available in BF16 and FP8 precision at launch, and community members estimate a full BF16 checkpoint needs about 4.9TB of storage, while a 1-bit quantized version from Unsloth is 397GB. The open-weight release lacks vision input and the 1M-token context length that appear in the official Qwen3.8-Max.