AI Model Release Tracker

Defensive powerhouse unlocked: GPT-5.4-Cyber, a fine-tuned GPT-5.4 variant, enables binary reverse-engineering for malware/vuln analysis without...

Meta's Llama 4 open-weight model is now public on Hugging Face, with claims of outperforming GPT-4o on several standard benchmarks – a major open-source push against proprietary giants.

OmniShow unifies text, ref images, audio, and pose for end-to-end human-object interaction video generation—the first to handle all four...

Key perspectives on NVIDIA's Ising launch, the first open-source AI models for quantum calibration and error correction:

• Performance edge: Ising...

Together AI's Introspective Diffusion LM is the first DLM to match autoregressive (AR) quality while outperforming prior DLMs in quality and efficiency, delivering 3x higher throughput than SOTA.

Emerging open multimodal video model flips the paradigm: extends diffusion-based video generator for unified generation & understanding, tackling...

Democratizing AI training: Counters massive models with just 8.7M parameters, trained on 60k synthetic fish chats.

• Easy access: Browser demo, Colab...

Gemini Robotics-ER 1.6 launches as DeepMind's reasoning-first model for robots, excelling in messy real environments over prior 1.5 and Gemini 3.0...

MCAT introduces scaling for many-to-many speech-to-text translation using MLLMs, advancing beyond Whisper's ASR and NLLB's MT capabilities.

Alibaba's flagship drops 78.8% SWE-bench Verified and 1M token context via hybrid linear attention + sparse MoE.

Key strengths:

• Agentic coding:...

New gold-standard eval: PsychiatryBench draws from expert psychiatric textbooks for 11 tasks (diagnosis, treatment, follow-up) with 5,188 items—fixing...

Agent-ready multimodal access without complex integrations: MMX-CLI exposes MiniMax's text/image/video/speech/music/vision/search via simple shell...

ByteDance releases OmniShow, an end-to-end framework unifying text, image, audio, and pose for high-quality human-object interaction video generation....

Quick experiment-ready release via tweet:

• arXiv abs: https://t.co/6dSiAlmzhw
• Repo: https://t.co/GQk9x6CyfW
• Model weights: https://t.co/V5oR5ajhIP

Open-Source Multimodal Releases

• 🔥 Audio Flamingo Next: NVIDIA and University of Maryland researchers released AF-Next, a fully open Large...

Open-source AF-Next advances multimodal audio understanding for speech, sounds, and music.

• Three variants: AF-Next-Instruct (QA), AF-Next-Think...

Real-world wins from testing Meta's new Muse Spark multimodal model powering Meta AI:

• Contemplating mode deploys multiple AI agents for reliable...

Emerging open Chinese powerhouse: Baidu's new ERNIE-Image-8B image model launches soon as an open model drawing attention.

• Tool-ready...

Elephant Alpha, a new stealth 100B instant model, heralds the Model Tsunami with more releases incoming. Its relatively small size means it won't rival giants, but signals accelerating AI proliferation.

• Zhipu AI's GLM-5.1 update delivers a 28% coding benchmark leap over GLM-5, hitting 94.6% of Claude Opus 4.6
• GLM-5 already topped open models on...