There is a flurry of AI model releases, all right after DeepSeek made their splash last week.
OpenAI announced the release of its new artificial intelligence model, called o3-mini, for free. The o3-mini model is reported to match its predecessor in areas like mathematics and science but operates at a lower cost and delivers faster responses. Users on OpenAI’s Pro package will receive unlimited access, while those on the Plus plan will experience increased limits compared to free users.
Google quietly launched its next-generation flagship AI model, Gemini 2.0 Pro Experimental, through a changelog for its Gemini chatbot app. Google claims it offers better factual accuracy and enhanced performance for coding and mathematics tasks. However, the model is still in early preview, meaning it may exhibit unexpected behaviors and errors.
Microsoft announced that its reasoning model, Think Deeper, developed by OpenAI, will now be available for free to all users of its Copilot software. This announcement comes just two weeks after Microsoft increased the cost of Copilot by three dollars per month. The new model, which is designed to help produce more accurate outputs by mimicking human reasoning through intermediary steps, aims to reduce errors and improve overall quality. The announcement has sparked some skepticism, as some users have reported a significant increase in their monthly subscription fees, raising questions about the true meaning of “free” in this context.
Mistral AI launched its Small 3 model, which the company claims is the most efficient in its category and optimized for quick responses. With 24 billion parameters, Small 3 is an open-source model designed for local use, capable of achieving over eighty-one percent accuracy on the MMLU benchmark test. The model has been tested with more than one thousand coding and general prompts by human evaluators, who preferred Small 3 over competitors like Gemma-2 and Qwen-2.5. Mistral recommends this model for applications such as customer-facing virtual assistants and fraud detection in financial services, stating that it can be fine-tuned to create highly accurate subject matter experts. The Small 3 model is available for use on devices like a MacBook with a minimum of thirty-two gigabytes of RAM. Mistral plans to release more models with enhanced reasoning capabilities in the near future.
And I wanted to observe that DeepSeek wasn’t done — DeepSeek launched a new family of multimodal AI models named Janus-Pro, which claim to outperform OpenAI’s DALL-E 3. The models range in size from one billion to seven billion parameters, with larger models generally demonstrating better problem-solving capabilities. Janus-Pro operates under an MIT license, allowing for unrestricted commercial use. Notably, the largest model, Janus-Pro-7B, has surpassed DALL-E 3 in performance on two benchmarks, GenEval and DPG-Bench.
Why do we care?
The models may, and should, move quickly to commoditization, and move the value to the application layer. This is continued good news to providers. At some point, tracking models will become entirely irrelevant. There’s value now, particularly to watch for capabilities early that can appear in products and be used in services.