In January, the release of open models took a bit of a dip compared to the groundbreaking year of 2025. While there were still some notable and impressive models introduced, the overarching sentiment in the AI community is one of anticipation for upcoming releases. Whispers of DeepSeek V4’s imminent launch, packed with remarkable features, have flooded the news cycle, hinting at a more competitive landscape for open models in the near future.
Across the broader AI ecosystem, discussions have been buzzing this weekend about the potential debut of Claude Sonnet 5, which could be just around the corner. This anticipation adds an exciting layer to what’s currently available—there’s plenty to explore with the new open models already at our fingertips!
Let’s take a closer look at some of the latest releases:
LFM2.5-1.2B-Instruct by LiquidAI: Continuing their pretraining journey from 10 trillion tokens to an impressive 28 trillion, LiquidAI has delivered an update that genuinely surprised us. In our evaluations, it was astonishingly close to the performance of the Qwen3 4B 2507 Instruct, a model we rely on daily, despite being over three times smaller! When pitted against the larger Qwen3 1.6B, we found ourselves favoring LFM2.5 almost every time. This round, LiquidAI also rolled out several variants simultaneously, including a Japanese version, a vision model, and an audio model.
Trinity-Large-Preview by arcee-ai: This model features an ultra-sparse Mixture of Experts (MoE) design, boasting a total of 400 billion parameters, with 13 billion active during operation. Developed by an American firm, they’ve also shared a technical report detailing insights and two foundational model variants—one prior to annealing and another following the pre-training processes. More technical information and the team’s motivations can be explored in our interview with the founders and pre-training lead.
Kimi-K2.5 by moonshotai: This model has undergone continuous pre-training on 15 trillion tokens and is notably multimodal! Users on Twitter have started substituting K2.5 for Claude 4.5 Opus when needing a less robust yet more budget-friendly alternative. However, the renowned writing capabilities of K2 and its successor seem to have diminished in favor of improved coding and agency-driven functions.
GLM-4.7-Flash by zai-org: A compact version of GLM-4.7, this model matches the small Qwen3 MoE in size, featuring 30 billion total parameters, with 3 billion of them active.
K2-Think-V2 by LLM360: Building upon their earlier models, this truly open reasoning model offers exciting advancements for users.
As we delve into this issue, it’s worth noting the exceptional quality of various niche small models across the ecosystem. From Optical Character Recognition (OCR) to embeddings and even song generation, this installment showcases a variety of models excelling in their respective areas. The challenge often lies in locating these gems, but they are indeed present, delivering remarkable capabilities to meet today's diverse needs.
