Model	techniques
DeepSeek-V4.1-Flash core	Filtering, deduplication, and difficulty calibration coreDeterministic pre-splitting of ultra-long documents usedDomain-specific multimodal datasets usedScaling-ladder-guided data and model planning usedUnion-based integration of text-only and multimodal corpora used—
NVIDIA-Nemotron-3-Ultra-550B-A55B used	Gaussian-based data mixture and curriculum construction usedTwo-phase curriculum used—
MiMo-V2.6-Flash core	Sample Mixer coreAgent-centric multimodal data mixture usedData Scheduler usedDynamic sampling for mixed-task reinforcement learning usedSmooth weighted round-robin source scheduling used—
MiMo-V2.5 used	Three-stage data mixture strategy used—
GLM-5.3 used	Vulnerability-discovery data inclusion used—
MiniMax-M3 used	Interleaved data usedInterleaved multimodal training used—
Gemma 4 31B used	Attribution, hedging, and refusal data subsets used—
Kimi K3 used	Long-context data upsampling used—
Laguna-S-2.1 core	AutoMixer coreBlind pairwise bucket-boundary calibration usedKL-regularized data-mixture optimization usedPass-rate-based RL task sampling usedPrior-constrained Dirichlet mixture exploration used—
MiMo-V2.5-Pro used	Three-stage data mixture strategy used—
MiMo-V2.6-Pro core	Sample Mixer coreAgent-centric multimodal data mixture usedData Scheduler usedDynamic sampling for mixed-task reinforcement learning usedSmooth weighted round-robin source scheduling used—
Qwen3.5-397B-A17B used	Three-stage pre-training curriculum used—
gpt-oss-120b used	Text-only pre-training corpus usedProtocolQA-aligned RL training datasets evaluated—

data mixture & curriculum

In this branch 26

filed here 26

By model