🔍 LLM DAILY

Your Daily Briefing on Large Language Models

December 17, 2025

HIGHLIGHTS

• Allen Institute for AI has launched Molmo 2, an 8B parameter multimodal model capable of analyzing both images and videos with advanced capabilities including precise pixel coordinate grounding, trained entirely on open datasets.

• Researchers have developed "Model-First Reasoning" (MFR), a novel approach that reduces LLM hallucinations by 60% during complex planning tasks by having models explicitly define the problem space before attempting solutions.

• Leona Health secured $14M in seed funding led by a16z to develop an AI co-pilot helping Latin American doctors manage patient messages on WhatsApp, addressing a critical healthcare communication challenge.

• Chai Discovery, an OpenAI-backed biotech firm, has reached unicorn status with a $130M Series B at a $1.3B valuation for their foundation models that predict molecular interactions for drug discovery.

• The community-driven "awesome-chatgpt-prompts" repository has surpassed 139,000 stars on GitHub, evolving from a simple README to a robust system for sharing and discovering effective prompts.

BUSINESS

Funding & Investment

• Leona Health: Raised $14M seed funding led by Andreessen Horowitz (a16z) to develop an AI co-pilot that helps Latin American doctors manage patient messages on WhatsApp. (2025-12-16)

• Chai Discovery: The OpenAI-backed biotech firm raised $130M Series B at a $1.3B valuation. The company is building foundation models for drug discovery, specifically to predict interactions between molecules for potential cures. (2025-12-15)

• Serval: Sequoia Capital announced a partnership with Serval, an enterprise AI automation company focused on empowering IT departments. (2025-12-11)

M&A and Partnerships

• Nvidia Acquires SchedMD: Nvidia has acquired SchedMD, the lead developer of Slurm, as part of its strategy to expand its open-source offerings. The acquisition coincides with Nvidia's launch of the Nemotron 3 family of open source AI models. (2025-12-15)

• Disney-OpenAI Deal: Disney's partnership with OpenAI is exclusive for just one year, after which Disney can sign similar deals with other AI companies. (2025-12-15)

Company Updates

• OpenAI Releases GPT Image 1.5: OpenAI has launched GPT Image 1.5 for ChatGPT, promising 4x faster generation, better instruction-following capabilities, and more precise edits. This release is seen as an escalation in its rivalry with Google Gemini. (2025-12-16)

• Meta's AI Glasses Update: Meta has added a new "conversation focus" feature to its AI glasses that uses the device's open-ear speakers to amplify the voice of the person you're talking to, improving hearing in noisy environments. (2025-12-16)

• DoorDash Launches Zesty: DoorDash has introduced Zesty, an AI-powered social app for discovering new restaurants. Users can view and share photos and comments about restaurants, discover content from others, and follow people like on traditional social networks. (2025-12-16)

• Google Tests Email Productivity Assistant: Google is testing a new email-based productivity assistant that will be available to AI Pro and Ultra plan users over 18 in North America. (2025-12-16)

Market Analysis

• Creative Commons AI Policy: Creative Commons has announced tentative support for AI 'pay-to-crawl' systems, offering support for the concept of an AI marketplace with suggested guiding principles. This represents an important development in the evolving landscape of AI content creation and rights management. (2025-12-15)

PRODUCTS

Allen Institute for AI Releases Molmo 2, a Multimodal Open Source Model

Company: Allen Institute for AI (Ai2) - Non-profit research lab
Released: 2025-12-16
Link: https://allenai.org/molmo

Ai2 has released Molmo 2, an impressive 8B parameter multimodal model capable of analyzing both images and videos. The model demonstrates advanced capabilities including video QA, counting and pointing within videos, dense captioning, and returning grounded answers with precise pixel coordinates and timestamps. What's particularly notable is that Molmo 2 was trained entirely with open datasets, maintaining Ai2's commitment to open-source AI development. The model appears to deliver strong performance despite its relatively modest parameter count compared to larger commercial multimodal models.

Allen Institute for AI Announces Olmo 3 Family of Language Models

Company: Allen Institute for AI (Ai2) - Non-profit research lab
Released: 2025-12-15
Link: https://www.reddit.com/r/LocalLLaMA/comments/1pniwfj/ai2_open_modeling_ama_ft_researchers_from_the/

Ai2 has also released Olmo 3, a family of fully open language models ranging from 7B to 32B parameters. The release includes Base, Instruct, and Thinking variants with long-context support. True to Ai2's open-source philosophy, the models come with complete transparency including open training recipes and checkpoints. This release represents a significant contribution to the open-source AI ecosystem, allowing researchers and developers to build upon truly open foundation models.

Cyreal: New JAX Dataloader Released for Machine Learning

Company: Independent developer (smorad)
Released: 2025-12-16
Link: https://github.com/smorad/cyreal

A new JAX dataloader called Cyreal has been released, addressing a common pain point in the JAX ecosystem. Marketed as fast, lightweight, and flexible, Cyreal aims to provide a native JAX solution for data loading rather than relying on PyTorch's dataloader, which Google's own documentation often recommends. While the developer notes that this is a new library likely containing bugs, it represents an important tool for the JAX community. Documentation is available at https://smorad.github.io/cyreal/cyreal.html.

TECHNOLOGY

Open Source Projects

awesome-chatgpt-prompts

A community-driven collection of ChatGPT prompts with 139,000+ stars. The project has evolved from a single README to a more robust system that allows users to share, discover, and collect effective prompts. All prompts are accessible via prompts.chat or in a structured CSV format on GitHub.

OpenAI Cookbook

Official examples and guides for using the OpenAI API with nearly 70,000 stars. Recently updated with new guidance for GPT-Image-1.5 prompting, the repository serves as a comprehensive reference for developers implementing OpenAI's technologies. Content is also accessible through a dedicated website at cookbook.openai.com.

Models & Datasets

Text-to-Image Models

Tongyi-MAI/Z-Image-Turbo

A high-performance text-to-image model with nearly 300K downloads and 2,800+ likes. Z-Image-Turbo implements advanced diffusion techniques and offers a custom ZImagePipeline for efficient image generation.

Voice and Speech Synthesis

microsoft/VibeVoice-Realtime-0.5B

A lightweight (0.5B parameters) real-time text-to-speech model developed by Microsoft. With 158K+ downloads, it specializes in streaming text input and long-form speech generation, built upon the Qwen2.5-0.5B architecture. The model enables responsive, continuous speech synthesis for interactive applications.

Multimodal Models

zai-org/AutoGLM-Phone-9B

A 9B parameter multimodal model optimized for mobile phone-based interactions. Built on the GLM-4.1V-9B-Base architecture, this model handles image-text-to-text tasks and functions as a conversational agent with 51K+ downloads.

zai-org/GLM-4.6V-Flash

A versatile multimodal model supporting both Chinese and English. This model handles any-to-any conversational tasks with strong image understanding capabilities and has accumulated over 100K downloads.

Datasets

Anthropic/AnthropicInterviewer

A dataset from Anthropic with 9,600+ downloads focused on interview-style interactions. The medium-sized collection (1K-10K samples) is formatted as CSV and designed for training conversational AI models to handle complex dialogue structures.

OpenMed/Medical-Reasoning-SFT-GPT-OSS-120B

A large-scale medical dataset (100K-1M samples) for training models on healthcare reasoning tasks. This resource is specifically designed for supervised fine-tuning of large language models for medical applications.

TuringEnterprises/Turing-Open-Reasoning

A specialized dataset for question-answering tasks across multiple scientific domains including chemistry, physics, math, biology, and code. Despite its smaller size (<1K samples), it has gained significant traction with nearly 15K downloads.

Interactive AI Spaces

ResembleAI/chatterbox-turbo-demo

A Gradio-based demo showcasing ResembleAI's conversational voice capabilities, allowing users to interact with an AI system using natural language and receiving voice responses.

Tongyi-MAI/Z-Image-Turbo

The official demo space for the Z-Image-Turbo model, enabling users to experience its text-to-image generation capabilities directly through a Gradio interface. This space has garnered over 1,400 likes, demonstrating significant user interest.

Wan-AI/Wan2.2-Animate

A highly popular animation generation space with over 2,700 likes. Wan2.2-Animate allows users to create animated content from static images or text prompts using the Wan2.2 model architecture.

HuggingFaceTB/smol-training-playbook

A research-focused Docker space with 2,600+ likes that provides a comprehensive playbook for training smaller, more efficient language models. It includes data visualization tools and follows a scientific paper format to present training methodologies.

mistralai/Ministral_3B_WebGPU

A static space demonstrating Mistral AI's 3B parameter model running directly in browsers using WebGPU technology. This implementation showcases how compact but capable language models can be deployed client-side without requiring server infrastructure.

RESEARCH

Paper of the Day

Model-First Reasoning LLM Agents: Reducing Hallucinations through Explicit Problem Modeling (2025-12-16)

Authors: Annu Rana, Gaurav Kumar

Institutions: [Not explicitly stated]

Why it's significant: This paper introduces a novel paradigm for LLM reasoning that addresses one of the most critical challenges in LLM agents—the tendency to hallucinate during complex planning tasks. By bringing classical AI planning principles to modern LLMs, it represents a significant advancement in developing more reliable AI systems.

Summary: The authors propose Model-First Reasoning (MFR), a two-phase approach where LLMs first construct an explicit model of the problem (defining entities, state variables, actions, and constraints) before solving it. Their experiments across diverse domains show MFR reduces constraint violations by 60% and improves solution consistency by 65% compared to Chain-of-Thought and ReAct approaches. This work bridges the gap between classical symbolic AI and modern neural approaches, creating more reliable and interpretable reasoning systems.

Notable Research

SparseSwaps: Tractable LLM Pruning Mask Refinement at Scale (2025-12-11)

Authors: Max Zimmer, Christophe Roux, Moritz Wagner, Deborah Hendrych, Sebastian Pokutta

This research introduces a novel approach to LLM pruning that optimizes pruning masks without expensive retraining, enabling significant parameter reductions while maintaining performance across Llama models of various sizes.

UniUGP: Unifying Understanding, Generation, and Planing For End-to-end Autonomous Driving (2025-12-10)

Authors: Hao Lu, Ziyang Liu, Guangfeng Jiang, et al.

The researchers present an innovative framework that combines understanding, generation, and planning capabilities in autonomous driving systems, leveraging specialized datasets and visual causal learning to handle complex scenarios that typically challenge traditional systems.

Reasoning-Style Poisoning of LLM Agents via Stealthy Style Transfer (2025-12-16)

Authors: Xingfu Zhou, Pengfei Wang

This paper reveals a concerning security vulnerability in LLM agents by demonstrating how attackers can manipulate an agent's reasoning style through subtle poisoning attacks, highlighting the need for robust runtime monitoring systems.

Cornserve: Efficiently Serving Any-to-Any Multimodal Models (2025-12-16)

Authors: Jeff J. Ma, Jae-Won Chung, Jisang Ahn, et al.

The authors introduce an efficient serving system for multimodal models that significantly reduces both memory requirements and request latency, addressing key challenges in deploying complex AI systems that handle multiple modalities simultaneously.

LOOKING AHEAD

As 2025 draws to a close, multimodal AI systems are reaching unprecedented levels of integration with physical environments. The Q4 breakthroughs in embodied AI suggest that by mid-2026, we'll see the first commercial robots with truly generalizable reasoning capabilities across physical tasks. Meanwhile, the regulatory landscape continues to evolve rapidly, with the EU's AI Harmonization Act set for implementation in Q1 2026 and similar frameworks emerging in Asia-Pacific markets.

Looking toward 2027, we anticipate the first "cognitive architecture" models that maintain persistent memory and goal-directed reasoning across extended timescales. These developments, coupled with the projected 70% reduction in inference costs, will likely accelerate AI adoption in healthcare and industrial automation while raising new questions about human-AI collaboration paradigms.