What’s New in Large Language Models? DeepSeek V3, Qwen2.5‑Omni, Gemini 2.5 Pro, and GPT‑4o Unpacked

DeepSeek‑V3 Update

DeepSeek released version DeepSeek‑V3‑0324 on March 25, a minor upgrade that retains the same base model while improving the post‑training pipeline. The open‑source checkpoint and tokenizer_config.json are updated. The model has roughly 660 B parameters, a context length of 128 K tokens for the open‑source version, and 64 K tokens via the web, app, and API interfaces.

Key capability upgrades include stronger reasoning performance, enhanced front‑end code generation (e.g., HTML), higher‑quality Chinese writing, and more accurate Chinese search results. The model also shows improvements in tool calls, role‑play, and general chat interactions.

Qwen2.5‑Omni‑7B Open‑Source Release

Alibaba’s Tongyi Qwen announced Qwen2.5‑Omni‑7B on March 27 as the first end‑to‑end multimodal large language model capable of processing text, images, audio, and video simultaneously. It employs a dual‑core Thinker‑Talker architecture with Position‑Embedding (TMRoPE) to fuse multimodal signals.

With only 7 B parameters, the model can be deployed on mobile devices and cloud services. It is available on the Magic‑Da community, Hugging Face, and can be tried directly in Qwen Chat. Features include real‑time text and natural‑voice generation, handling 10‑20 objects per scene, and style conversion from sketches to photorealistic images.

Google Gemini 2.5 Pro

Google unveiled Gemini 2.5 Pro on March 26, succeeding Gemini 2.0 Flash‑Thinking. The model offers a 1 M‑token context window and native multimodal support for text, audio, images, video, and code.

Benchmarks show strong performance on reasoning, mathematics, scientific, and programming tasks, comparable to Claude 3.7 Sonnet and Grok 3. Gemini 2.5 Pro is currently accessible to Gemini Advanced paid users via Google AI Studio and will be rolled out to Vertex AI.

OpenAI GPT‑4o

OpenAI released GPT‑4o on March 26, adding native multimodal image generation to ChatGPT. The model can embed text within images, maintain context across interactions, and bind up to 10‑20 objects in a single scene.

Pricing is reduced to $0.10 per M input tokens (≈10× cheaper than GPT‑4 Turbo) and $0.15 per M output tokens. GPT‑4o is available in ChatGPT Plus, Pro, Team, and Free tiers, with API access slated for future release. Users can specify detailed image creation parameters such as aspect ratio, hex color codes, and transparent backgrounds; rendering may take up to one minute.