Kuaishou Keye-VL-2.0 in-depth review: 30B open source video model, 256K ultra-long context fighting Gemini

Kuaishou is officially open sourceKwai Keye-VL-2.0-30B-A3B——A hybrid expert (MoE) multi-modal basic model with 30B parameters. Only 3B parameters are activated during operation.DeepSeekSparse Attention (DSA)The technology achieves 256K lossless ultra-long context processing and can directly parse hours of complete video.

core competencies

The technical architecture of Keye-VL-2.0 is built around three core breakthroughs:

• DeepSeekSparse Attention (DSA): Keye-VL-2.0 is the first production-level visual language model to integrate DSA into the GQA architecture.
• Native resolution visual encoding: Adopt native resolution encoding strategy to avoid the loss of details caused by traditional scaling and ensure accurate spatio-temporal positioning capabilities in high-resolution videos.
• Cross-modal multi-teacher online policy distillation (MOPD): Through multi-teacher knowledge distillation + reinforcement learning, it injects multi-task capabilities while preventing catastrophic forgetting, allowing the model to perform well in high-order agent capabilities such as code writing, tool calling, and web search.

Benchmark performance

In the five video understanding benchmarks, Keye-VL-2.0 takes the lead among models of the same scale, and even challenges the closed-source giants on some indicators:

• LongVideoBench: Ranked first, leading the industry in long video understanding ability
• Video-MME-v2: Ranked first, with the best performance in multi-modal video Q&A
• TimeLens: Fine-grained time positioning capability surpasses Gemini 3 Flash
• mathematical reasoning: Remain competitive among visual language models of the same scale

It is worth emphasizing that, with 30B total parameters and only 3B activation parameters, Keye-VL-2.0 not only surpasses 200B+ open source models (such as Qwen3-VL-235B) in time understanding capabilities, but also demonstrates excellent agent collaboration capabilities in multiple rounds of tool interaction and status tracking tests.

Technical Highlights: Four-Phase Pre-Training Course

Keye-VL-2.0 adopts a unique four-stage pre-training strategy: from basic visual language alignment, to large-scale multi-modal data injection, to sparse attention special training, and finally to logical trajectory strategy optimization through reinforcement learning (Context-RL + Video-RL).

User experience/limitations

Advantages: Apache 2.0 open source protocol, fully open for commercial use; 30B MoE architecture with only 3B activation parameters, extremely low inference cost; adapted to mainstream frameworks such as Transformers, vLLM, SGLang, Docker Model Runner, etc.; natively equipped with Agent capabilities (search, tool invocation, code writing); open sourced simultaneously on Hugging Face and GitHub, the community ecosystem can be expected.

limitations: As a video understanding specialty model, it is not as good as a general-purpose large model of the same scale in pure text tasks; the 30B parameter level requires certain GPU resources (although there are fewer activation parameters, the total parameter amount still requires corresponding video memory); Kuaishou's influence in the domestic open source community is still being established compared to Alibaba, Deep Search, etc.; Agent capabilities (code, tools, search) are still in the first generation, and there is a gap with specialized programming agents.

Overall Score

Overall rating: 8.4/10