Which of this branch's techniques each model's own documents describe, and how strongly. Under each model: its strongest treatment anywhere in the branch.

Model	filed here	inference engine	training framework	kernel & quantization library	agent product	infrastructure service
GLM-5.3-Flash used	—	—	—	—	—	Automatic provider failover used—
DeepSeek-V4.1-Flash core	—	—	—	deepseek-recipe optional—	Claude Agent SDK with version-specific native tool interfaces used—	DeepSeek Elastic Compute (DSec) coreCustom placement engine usedDurable peer mailbox used—
Hy4-preview used	—	vLLM or SGLang serving used—	—	AngelSlim usedFLASHMLA_SPARSE used—	—	—
NVIDIA-Nemotron-3-Ultra-550B-A55B used	CUDA usedData Designer used—	Environment variable propagation to subprocesses usedvLLM usedSGLang optionalTensorRT-LLM optional—	Hugging Face Transformers usedMegatron-LM usedNeMo usedNeMo Gym used—	Container-baked precompiled artifacts usedFlashInfer usedFlashInfer NVLinkOneSided usedNVIDIA Model-Optimizer usedCUTEDSL optionalCUTLASS optional—	Droid usedOpenCode usedOpenHands usedStirrup used—	Dependency unification usedFresh-clone repository sanitization usedLocal container-image caching usedPer-node pooling of initialization actors usedReplacing short-lived Ray actors with tasks usedRay optional—
MiMo-V2.6-Flash used	—	vLLM tensor parallelism used—	—	—	—	—
DeepSeek-V4-Flash used	—	—	—	—	—	Write-ahead log usedZ3 SMT solver integration used—
MiMo-V2.5 optional	—	—	—	—	Claude Code optionalKilo optionalOpenCode optional—	—
GLM-5.3 used	—	—	—	AITER used—	—	—
Hy3 mentioned	—	—	—	AngelSlim mentioned—	—	—
GLM-5.2 used	—	KTransformers optionalSGLang optionalUnsloth optionalvLLM optionalvLLM-Ascend optionalxLLM optional—	Hugging Face Transformers optional—	—	Claude Code usedOpenHands used—	—
MiniMax-M3 used	—	KTransformers optionalSGLang optionalUnsloth optionalvLLM optional—	Hugging Face Transformers optional—	flash-moba evaluatedFlash-Sparse-Attention evaluated—	Claude Code usedMini-SWE-Agent usedOpenCode usedPi used—	—
DeepSeek-V4-Flash-Vision-Exp default	Separate encoding and inference modules default—	vLLM optional—	—	—	—	—
DeepSeek-V4-Pro used	—	—	—	—	—	Write-ahead log usedZ3 SMT solver integration used—
DeepSeek-V4-Pro-0813 optional	—	—	—	—	—	Automatic provider failover optional—
Gemma 4 31B core	—	litert-lm serve used—	—	—	—	Automatic provider failover core—
Inkling used	—	—	—	—	—	Automatic provider failover used—
Kimi K3 core	—	—	—	MiniTriton usedMiniTriton dual-mode tensor library usedTriton mentioned—	OpenClaw used—	Incremental sandbox checkpointing and resumption coreSandbox pause and resume coreMooncake used—
Laguna-S-2.1 core	Shared composable codebase for research and production used—	Atlas inference library used—	DAG-based asset lineage tracking coreExperiments as Code core—	—	—	End-to-end lineage coreHTTP/HTTPS sidecar proxy usedIn-house batch scheduler usedModel Factory usedVendoring external dependencies used—
MiMo-V2.5-Pro optional	—	—	—	—	Claude Code optionalKilo optionalOpenCode optional—	—
MiMo-V2.6-Pro used	—	vLLM tensor parallelism usedSGLang optionalvLLM optional—	—	—	—	—
NVIDIA-Nemotron-3.5-Lightning-30B-A3B used	Nemotron Parse usedModel routing and orchestration optional—	TensorRT-LLM used—	—	—	—	NeMo Guardrails usedNeMo Switchyard used—
Qwen3.5-397B-A17B used	—	vLLM language-model-only mode optional—	—	—	Qwen-Agent used—	Automatic provider failover used—
Qwen3.6-35B-A3B core	fastsafetensors used—	Dedicated serving engines optional—	—	Marlin MoE backend used—	—	Automatic provider failover core—
Qwen3.8-Flash-Next used	—	—	—	FlashQLA used—	—	—
Step-3.7-Flash used	—	llama.cpp used—	—	FlashInfer TensorRT-LLM backend usedFlashInfer optional—	—	—