Latest open artifacts (#17): NVIDIA, Arcee, Minimax, DeepSeek, Z.ai and others close an eventful year on a high note

2026-01-05 · Source: Interconnects AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Advanced, quick

Summary

The open ecosystem continues its rapid expansion into 2026, with several significant model releases. LLM360, a project from MBZUAI, introduced K2-V2, a 70B dense model with fully open pre-training (12T tokens) and SFT data, generated using GPT-OSS 120B. NVIDIA updated its Nemotron series with NVIDIA-Nemotron-3-Nano-30B-A3B-BF16, featuring a Mamba2-Transformer MoE architecture and plans for larger 100B and 500B models in H1 2026 utilizing Latent MoE and multi-token prediction. Arcee-ai released Trinity-Mini (26B-A3B MoE) and Trinity-Nano (6B-A1B MoE), with a 420B-A13B MoE Large model planned. Zhipu's GLM-4.7, released before its January 8th IPO, demonstrates strong performance in tasks like GPVal-AA and DesignArena, despite not leading on academic benchmarks. Other notable releases include Llama-3.3-8B-Instruct, ServiceNow-AI's Apriel-1.6-15b-Thinker, XiaomiMiMo's 309B-A15B MoE MiMo-V2-Flash, and DeepSeek-V3.2, including a "Speciale" version claiming gold-medal performance on IMO and IOI.

Key takeaway

For AI Architects evaluating new open-source models, prioritize those offering full data transparency like LLM360's K2-V2 or advanced architectures such as NVIDIA's Nemotron-3 MoE series. Your decision should weigh specific task performance, as seen with GLM-4.7 excelling in practical applications over academic benchmarks, against resource requirements like GLM-4.7's slower inference. Consider the upcoming larger models from NVIDIA and Arcee for future scaling needs.

Key insights

The open model ecosystem is rapidly evolving with diverse architectures, data transparency, and specialized performance.

Principles

• Full data transparency accelerates open model development.
• MoE architectures enhance scalability and performance.
• Specialized benchmarks reveal model strengths beyond SOTA.

Method

LLM360 generated SFT data using GPT-OSS 120B across three reasoning levels. ServiceNow-AI used GSPO with length and verbosity penalties to reduce token usage.

In practice

• Explore GLM-4.7 for UI generation tasks.
• Consider Arcee's Trinity-Mini for MoE capabilities.
• Utilize LLM360's K2-V2 for fully open-source training data.

Topics

• Open-source LLMs
• Mixture-of-Experts
• Model Architectures
• LLM Training Data
• AI Model Benchmarks

Best for: AI Architect, NLP Engineer, AI Scientist, AI Engineer, Machine Learning Engineer, AI Researcher

Related on AIssential

• Latest open artifacts (#18): Arcee's 400B MoE, LiquidAI's underrated 1B model, new Kimi, and anticipation of a busy month — Smaller open models are achieving competitive performance against larger counterparts through extensive pretraining and architectural innovations.
• MiniMax M2.7 Advances Scalable Agentic Workflows on NVIDIA Platforms for Complex AI Applications — MiniMax M2.7, a 230B-parameter MoE model, offers high efficiency and performance for agentic tasks via NVIDIA optimizations.
• Latest open artifacts (#22): Zyphra, Cohere, and Poolside are expanding the breadth of the ecosystem — The open model ecosystem is diversifying with varied actors and motivations, fostering innovation and specialized model development.
• Z.ai GLM-5: New SOTA Open Weights LLM — Chinese AI labs are rapidly advancing open-weight models, emphasizing cost-efficiency, long-context, and agentic capabilities.
• Nvidia’s Open Salvo, OpenAI’s Amazon Deal, Grok Cuts Video Prices, Recursive Language Models — Nvidia's Nemotron 3 Super offers a fast, open-source LLM for agents, leveraging hybrid architecture and co-designed hardware-software optimization.
• The KV-Cache of Small MoEs: Qwen3, Qwen3.5, GLM 4.7 Flash, and Nemotron 3 Nano Compared — MoE models achieve inference efficiency through diverse attention architectures and KV cache optimization strategies.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Interconnects AI.