2.5-24
An open-source large language model based on a mixture-of-experts architecture, with 24 billion total parameters and 2.45 billion activated parameters. It is designed specifically for local agent and tool-calling scenarios.
Timeline (last 90 days)
An AI company with MIT background (referred to in the text as "Council") released the open-source model 2.5-24, with 24 billion total parameters, activating only 2.45 billion parameters per token generated.
The four-quantized version of 2.5-24 is about 13.5G in size and can run on a laptop with 32G of memory; the 6 version is about 18.3G and requires 48G of memory.
In the 16 benchmarks officially published for 2.5-24, it scored 27.69 on automated tasks, with second place at 8.72, leading by about 3.2 times.
2.5-24 scored 48.09 in actual agent testing, compared with second place's 38.19.
2.5-24 scored 80.71 in the tool sandbox test, ranking first.
2.5-24 took first place in all three instruction-following tests.
The generation speed officially given for 2.5-24 is 184 characters per second.
On the first day of release, 2.5-24 provided model files for Apple chips, cross-platform, and server-side corresponding versions, with broad ecosystem support.
The 2.5-24 weights can be downloaded for free, with no secondary fees for local calls; the cloud API input price is about on the order of one-tenth of a US dollar per million tokens.
2.5-24 does not have an advantage in hardcore science Q&A, complex code, or fully autonomous web operations, and is positioned as a dedicated player for local agents.
1 more timeline events
All Facts (11)
Running 2.5-24 locally can ensure that data such as emails, files, schedules, and code read by the agent does not leave the machine, improving privacy and security.
50%UnverifiedThe four-quantized version of 2.5-24 is about 13.5G in size and can run on a laptop with 32G of memory; the 6 version is about 18.3G and requires 48G of memory.
50%UnverifiedIn the 16 benchmarks officially published for 2.5-24, it scored 27.69 on automated tasks, with second place at 8.72, leading by about 3.2 times.
50%Unverified2.5-24 scored 48.09 in actual agent testing, compared with second place's 38.19.
50%Unverified2.5-24 scored 80.71 in the tool sandbox test, ranking first.
50%Unverified2.5-24 took first place in all three instruction-following tests.
50%UnverifiedThe generation speed officially given for 2.5-24 is 184 characters per second.
50%UnverifiedThe 2.5-24 weights can be downloaded for free, with no secondary fees for local calls; the cloud API input price is about on the order of one-tenth of a US dollar per million tokens.
50%Unverified2.5-24 does not have an advantage in hardcore science Q&A, complex code, or fully autonomous web operations, and is positioned as a dedicated player for local agents.
50%UnverifiedOn the first day of release, 2.5-24 provided model files for Apple chips, cross-platform, and server-side corresponding versions, with broad ecosystem support.
50%UnverifiedAn AI company with MIT background (referred to in the text as "Council") released the open-source model 2.5-24, with 24 billion total parameters, activating only 2.45 billion parameters per token generated.
50%