AI's Rapid Evolution: OpenAI's Garlic, Apple's Clara, Real-Time Voice & Avatars

The New Frontier: OpenAI's Garlic, Apple's Clara, Real-Time Voice & Avatars Reshape AI

A recent wave of breakthroughs has swept across the artificial intelligence landscape, showcasing remarkable advancements from leading tech giants. OpenAI declared an internal "code red" and revealed a new model, Garlic, while Apple introduced Clara, a novel approach to document compression. Microsoft unveiled a real-time text-to-speech (TTS) model, Vibe Voice, effectively eliminating awkward delays in AI conversations. Alibaba presented Live Avatar, an avatar system capable of indefinite video streaming, and Tencent released Huan Video 1.5, a video generator accessible to a broader audience. These developments, occurring almost simultaneously, underscore a period of intense innovation and competition in the AI domain.

OpenAI's "Code Red" and the Emergence of Garlic

The AI landscape recently saw an internal "code red" declared within OpenAI, triggered by Google's Gemini 3 reaching the top of LLM Marina charts. This declaration by Sam Altman indicated significant pressure and that competition had closed the gap, necessitating an emergency response.

In this context, OpenAI has been secretly developing Garlic, a new model. Mark Chen, Chief Research Officer, confirmed Garlic's strong performance in internal tests, where it reportedly beats Gemini 3 and Anthropic's Opus 4.5 in reasoning and coding. These two models had previously set the gold standard in these categories.

Garlic's innovation stems from a rebuilt pre-training system. Instead of forcing the model to absorb all fine-grain details initially, the process now focuses on broader connections first, then narrows into specifics. This shift is described as major internally and allows packing more knowledge into smaller models, which are cheaper to train and faster to run. The rise of capable small models from Mistral, Deepseek, and Chinese labs influenced OpenAI's response.

Garlic is separate from Shallot Pete, another internal project aimed at fixing pre-training bugs, indicating OpenAI is running multiple model lines to leapfrog competition. While no official launch date, Mark Chen stated "as soon as possible," with early next year being a safe bet. The work on Garlic has already enabled progress on the next major model.

In contrast, Anthropic, according to Dario Amodei, does not compete for the same audience as OpenAI and Google, focusing on enterprise customers. Their Claude code system achieved a $1 billion revenue run rate within six months of launch.

Apple's Clara: Rethinking Document Compression for AI

Apple quietly released Clara, a system designed to improve how AI handles long documents. Traditional chatbot systems often struggle with extensive texts, using methods that are slow, expensive, and less effective as documents grow. Clara addresses this by compressing documents into tiny memory tokens, which are super-dense summaries retaining important meaning. These tokens are then used for both retrieval and generation, allowing an AI to pull small bundles of compressed tokens instead of thousands of words.

Clara's innovative aspect includes training its retriever and generator together, allowing them to learn from each other, unlike most Retrieval Augmented Generation (RAG) systems that treat them separately.

The training involved approximately 2 million Wikipedia passages from 2021. A local Quen 32B model generated simple Q&A, multihop Q&A, and paraphrases. A verification loop of up to 10 rounds ensured data consistency. Two losses were used for training: cross entropy, to teach question answering with compressed memory, and MSE loss, to align memory and full document tokens. These details contribute to Clara's performance at high compression levels.

Clara's performance metrics are significant. At four times compression, its Compressor achieved an average F1 score of 39.86 on Natural Questions and Hot Pot QA, which is 5.37 points better than LLM Lingua 2 and over a point ahead of Pisco. In an Oracle setup, the model reached 66.76 F1, widely beating other methods. Compressed representations sometimes even outperformed full text retrieval pipelines like BGE plus Mistral 7B, delivering better results from shorter inputs. As a reranker, Clara achieved 96.21 recall at 5 on Hotpot QA, surpassing supervised retrievers.

Apple released three versions of Clara (Base, Instruct, and E2E) and the entire training pipeline, suggesting a potential larger push into the LLM space.

Microsoft's Vibe Voice: The End of AI Speaking Delays

Microsoft introduced Vibe Voice Realtime 0.5B, a real-time voice model that resolves the awkward pauses in AI assistants. This model can begin speaking in about 300 milliseconds, essentially instantly.

It is designed for agents and LLMs that speak while thinking. As the language model generates text, Vibe Voice simultaneously converts tokens into speech, allowing the LLM to continue producing the rest.

The model uses an acoustic tokenizer running at 7.5 hertz, based on a sigma VAE system with seven transformer layers and a 3,200 times downsampling factor from 24 kHz audio. A small four-layer diffusion head sits on top, conditioned on hidden states from a Quen 2.5 to 0.5B model. The total stack is around 1 billion parameters.

Performance on Libre Speech Test Clean shows a 2% word error rate and a speaker similarity score of 0.695. Vibe Voice is optimized for long-form speech, maintaining stability across conversations, and can generate about 10 minutes of audio within its 8K context window. It keeps output clean for assistant-style use cases by not attempting music or background noise. Microsoft recommends running it as a small microservice alongside an LLM, with perfect synchronization between text and audio streams.

Alibaba's Live Avatar: Infinite, Real-time Streaming Video

Alibaba, with major Chinese universities, released Live Avatar, an avatar system aiming to make animated avatars a real product. This system uses a 14 billion parameter diffusion model to generate video at over 20 frames per second in real time.

Users can speak into a mic, and the avatar responds instantly with smooth facial motion, gestures, and expressions. It supports streaming for over 10,000 seconds without drifting or losing identity.

This was achieved using several techniques. Distribution matching distillation shrunk a heavy multi-step video diffusion system into a model working with only four sampling steps. Timestep forcing pipeline parallelism spread the denoising process across GPUs for near-linear speedups, yielding an 84 times improvement.

The key innovation is solving the long video decay problem, where traditional auto-regressive video generators degrade over time (color shifts, face warping, identity drift). Live Avatar fixes this with three ideas:

Rolling row PE: Maintains stable positional information as video length increases.
Adaptive attention sync: Places the original reference frame with a generated one to prevent drift from the real distribution.
History corrupt: Adds controlled noise to the cache during training, teaching the model to recover from small mistakes.
Together, these elements allow Live Avatar to produce effectively infinite length streaming video without quality loss, a capability not seen before. Demos of two AI agents conversing with real-time avatars resemble sci-fi.

Tencent's Huan Video 1.5: Accessible High-Quality Video Generation

Tencent introduced Huan Video 1.5, described as one of the most accessible, high-quality video generators this year. With only 8.3 billion parameters, a small size for video systems, it delivers top-tier video quality. This includes smooth motion, strong prompt following, clean text rendering, and stable camera movement. A key feature is its operational speed.

Conclusion

The recent cascade of advancements from OpenAI, Apple, Microsoft, Alibaba, and Tencent highlights a vibrant and intensely competitive era in artificial intelligence. From OpenAI's urgent internal "code red" driving the development of the high-performing Garlic model to Apple's innovative Clara for efficient document compression, Microsoft's Vibe Voice offering instantaneous AI speech, Alibaba's Live Avatar enabling infinite real-time video streaming, and Tencent's accessible Huan Video 1.5, these breakthroughs collectively push the boundaries of what AI can achieve. Each development, in its unique domain, signifies a substantial leap forward, promising to reshape how we interact with technology and process information in the coming years.