{"id":56534,"date":"2026-09-01T11:35:09","date_gmt":"2026-09-01T03:35:09","guid":{"rendered":"https:\/\/www.1ai.net\/?p=56534"},"modified":"2026-09-01T11:35:09","modified_gmt":"2026-09-01T03:35:09","slug":"%e7%ab%af%e4%be%a7%e5%94%af%e4%b8%80%e7%99%be%e4%b8%87-token-%e4%b8%8a%e4%b8%8b%e6%96%87%e6%a8%a1%e5%9e%8b%e5%bc%80%e6%ba%90%ef%bc%9a%e7%a7%91%e5%a4%a7%e8%ae%af%e9%a3%9e%e8%af%8d%e5%85%83%e6%98%9f","status":"publish","type":"post","link":"https:\/\/www.1ai.net\/en\/56534.html","title":{"rendered":"The only one million Token context model open source on the side: the science big-tempo meta-fire X2.5-4B&1.7B officially released"},"content":{"rendered":"

September 1 News.iFLYTEK<\/a>The word \"Small Fire\" was announced todayOpen Source<\/a>STARFIRE X2.5-4B AND STARFIRE X2.5-1.7B BOTH END-SIDE GENERICLarge Model<\/a>.<\/p>\n

\"1xa005bd000v90hkp\"<\/p>\n

It was described that the two models had original support of up to 1 million Token context windows, the only product in the current end-side model that supported that level of context\u3002<\/p>\n

The two models use a mix of attention structure and optimize around core competencies such as intelligence, code, mathematics and command compliance. Completion of full-process training based on the National Computing Platform using approximately 20 trillion Token diversity data\u3002<\/p>\n

For example, after the sale of the product, the user can import a complete post-sale manual into a model which, in response to such complex issues as \u201cthe purchase of equipment after 10 days and the use of third-party consumables\u201d, is able to provide a comprehensive judgement across chapters and to provide ongoing advice under new conditions\u3002<\/p>\n

In the personal office scene, Starfire X2.5-4B completes the entire process from raw data analysis to bilingual report delivery. On the Domux Smart Home Test Set, the Starfire X2.5-1.7B control command end-to-end performance is 903% with an average response time of 0.85 seconds. Models also apply to robotic operational controls, target tracking, navigational decision-making and can be deployed in robotics or peripherals\u3002<\/p>\n

Starfire X2.5-4B and 1.7B support hardware platforms such as Weeda, Huacheng, Sealight and Backmoor, compatible with the mainstream reasoning frameworks of vLM, SSGLang, llama.cpp, which can be rapidly deployed through tools such as Ollama, LM Studio. As of this date, model weights have been opened on platforms such as Hugging Face, GitHub, and API has been on-line with the Stars Maas platform, free of charge for a limited time\u3002<\/p>\n

ON 7 SEPTEMBER, TIX WILL ALSO OFFICIALLY RELEASE THE STARFIRE X2.5-293B GENERIC LARGE MODEL, FURTHER UPGRADING CORE CAPABILITIES SUCH AS CODE AND INTELLIGENCE\u3002<\/p>\n

References:<\/p>\n