Unverified50% confidenceFactExact time
Compared with the previous generation LFM2-24B-A2B, LFM2.5-24B-A2B's architecture is unchanged (40 layers, 64 experts, top-4, convolution/GQA hybrid), with the main upgrades in training scale, long context, and reasoning capability: context increased from 32K to 131K, vocabulary from 65K to 128K, pre-training tokens from 17T to 33T, plus a newly added always-on chain of thought and 5-stage post-training (ending with agentic RL).
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedQwen 2.5 Coder, with only 32B parameters, matched the larger general-purpose model GPT-4o, indicating that code-specialized training has a punch-above-its-weight effect on complex array shape and broadcasting errors.70% similarUnverifiedThe author's publicly shared standard training configuration is: Krea 2 Standard model, Standard LoRA, Network rank 32, Alpha 32, Epochs 64 (saving every 4 epochs), Batch size 1, target resolution 0.25 MP, EMA 0.98.68% similarUnverifiedThe author's fine-tuning used the QLoRA method, with LoRA rank/alpha of 32/64, a learning rate of 1e-4, batch size 1 with gradient accumulation of 16, AdamW 8-bit optimizer, and training for 1 epoch.66% similarUnverifiedTraining simultaneously across 4 harnesses improved Liquid AI's LFM2.5-2.6B model score from 42% to 54%.65% similarUnverified千问3.8 27B基于千问3.5的整套架构体系，架构层面与3.5完全一致，仅做了继续预训练和后训练（含SFT与RL强化）65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1029968API
curl https://kongchang.com/api/v1/knowledge/claims/1029968MCP
get_claim(id=1029968)