Deep Chic unveiled the most recent update version of the open source reasoning model ‘R1’ and released detailed technologies in at some point. In complex reasoning, it has the same level of performance as ‘O3’ of Open AI and Google’s ‘Geminai 2.5 Pro’.
Deep Chic introduced the small print of ‘Deep Chic-R1-0528’, which improved performance than the present R1 model on the twenty ninth (local time).
The R1-0528 version said that it would provide greatly enhanced performance in high-level reasoning tasks resembling mathematics, science, business, and programming.
Like its predecessor, it would be released under the MIT license, business use and customization are freely allowed. Open source model weight HubIt’s distributed through, and an in depth document is provided for local distribution and API integration.
Existing deep API users are robotically updated to R1-0528 at no additional cost. The API usage fee is $ 0.14 per 1 million tokens, $ 0.035 for the discount time, and $ 2.19 per million tokens. Personal users Official websiteYou possibly can experience it without spending a dime after authentication of your cell phone or Google account.
Probably the most emphasized is that the output of hallucinations or misunderstandings decreased by about 45-50% in situations resembling writing or summary.
Along with improving performance, functional improvements for developers and researchers are noticeable. The addition of JSON output and performance call functions has made it easier to integrate the appliance, and the front end interaction has also improved easily and efficiently. As well as, the ‘Considering Mode’, which had been used before using a particular token, was modified to the system prompt method, making it easier to distribute.
Deep Chic said it has improved performance by optimizing algorithms within the post -processing stage and expansion of computing resources.
The benchmark said that it has risen significantly than the present version, which has caught up with advanced frontier models resembling Open AI’s O3 and Google’s ‘Geminai 2.5 Pro’.
‘AIME 2025’, which tests mathematics, said the accuracy has risen from 70%to 87.5%, and the typical variety of tokens used is increased from 12,000 to 23,000, showing a deep pondering process.
Within the coding test, ‘LiveCodebench’, the performance has doubled from 63.5%to 73.3%, and eight.5%to 17.7%in ‘Humanity’s Last Exam’, which is the best difficulty.
This has reached a competitive level with O3 and Geminai 2.5 Pro. After the R1 was released in January, Open AI and Google stressed that it surpassed it one after one other.
On at the present time, Deep Chic also unveiled the distilled model ‘Deep Chic-R1-0528-Q1 3-8B’, which used the text generated by R1-0528.
The model has a ten% higher performance than the Q13-8B, and has achieved a level much like ‘Q13-235B-Sinking’. Particularly, it is predicted to be executed in a single GPU with 16GB VRAM, which is predicted to be less hardware burden and useful in academic and industrial sites.
Aside from performance improvement, nonetheless, concerns were raised about sensitive questions related to freedom of expression.
The anonymous developer “XLR8HARDER”, which operates the AI response comparison platform speech map, tested the Deep Chic-R1-0528 and insisted that the Chinese government is more sensitive to the subject that the Chinese government is more sensitive than the previous version. “The R1-0528 is probably the most suppressed model of the Chinese government among the many deep-chic models which have been released thus far,” he said.
Some experts predicted that the R1-0528, which was released this time, could possibly be the last model of the R1 series, and Deep Chic will soon introduce the next-generation model ‘R2’ with more advanced performance.
By Park Chan, reporter cpark@aitimes.com
