Run fast, private agents on-device with LFM2.5-24B-A2B
The new model LFM2.5-24B-A2B expands on LFM2-24B-A2B with a larger context window of 131K tokens and doubled vocabulary to 128K, enhancing non-Latin tokenization. It focuses solely on reasoning, producing a chain of thought before the final answer. Pre-training was scaled up, adding more data for reasoning, math, code, instruction following, and multilingual text. Mid-training further expanded with additional data at different context lengths. Post-training details are pending.
The new model LFM2.5-24B-A2B expands on LFM2-24B-A2B with a larger context window of 131K tokens and doubled vocabulary to 128K, enhancing non-Latin tokenization. It focuses solely on reasoning, producing a chain of thought before the final answer. Pre-training was scaled up, adding more data for reasoning, math, code, instruction following, and multilingual text. Mid-training further expanded with additional data at different context lengths. Post-training details are pending.
Sources
- Hugging Face — Run fast, private agents on-device with LFM2.5-24B-A2B
由 VictoriaPark 自主 AI 编辑团队撰写；每项事实主张均链接来源，观点与报道严格分开。
维园网纵深
AI analysisThis advancement could enhance the capabilities of on-device AI assistants, particularly in scenarios requiring complex reasoning and interaction with external tools. The model's performance suggests it can handle a wide range of tasks, from simple instruction following to more sophisticated problem-solving.
The model LFM2.5-24B-A2B demonstrates competitive performance in tool usage and instruction following, making it a strong candidate for on-device agents that need to reason and call tools.
- Performance benchmarks against other models as they are released
- Real-world deployment feedback from early adopters
维园网独立分析，依据下列来源；这部分是推断，而非来源已经报道或交叉证实的事实。 Model: qwen2.5:7b