Mellum2 Technical Report
Several new AI models and challenges have been announced in recent research papers, advancing various aspects of artificial intelligence.
Mellum 2, a general-purpose language model specialized in software engineering, has been released. It is the successor to the completion-focused 4B dense Mellum model and features a 128K context window via a layer-selective YaRN[1]. Mellum 2 is competitive with open-weight baselines in the 4B-14B range while running at the per-token compute of a 2.5B dense model. The AgentDS competition, which consisted of 17 challenges across six industries, was participated in by 29 teams and 80 participants, highlighting the struggles of current AI agents with domain-specific data science tasks[2]. Human-AI collaboration outperformed AI-only baselines in the competition. Zamba2-VL, built on the hybrid language-model architecture Zamba2, is competitive with leading Transformer-based open-weight VLMs of comparable scale and substantially outperforms prior SSM-based and hybrid VLMs[3]. MOSS-Audio, a unified audio-language model, supports audio captioning, time-aware question answering, and timestamped transcription, achieving strong performance across general audio understanding and speech captioning tasks[4]. The CASTLE Challenge @ EgoVis 2026 will evaluate long-form egocentric video question answering over 600+ hours of multi-perspective recordings, requiring evidence from various sources including videos, transcripts, and auxiliary photos[5].
research-papermodel-releaseinfrastructureapplication
Background sources we checked (1)
- arxiv.org ↗ We present Mellum 2, an open-weight 12B-parameter Mixture-of-Experts (MoE) language model with 2.5B active parameters per token. Mellum 2 is a general-purpose language model specialized in software engineering, spanning code generation and editing, debugging, multi-step reasoning…
Sources cited (5)
- arxiv.org ↗ E
- arxiv.org ↗ E
- arxiv.org ↗ E
- arxiv.org ↗ E
- arxiv.org ↗ E