TELE-AI / DENSE THINKING / 2026

TeleChat3-36B-Thinking

TOTAL36.24Bparameters

ACTIVE36.24Bper token

DEPTH64repeated blocks

HIDDEN6,144main trunk

CONTEXT32Knative tokens

CHECKPOINT72.48GB20 shards · TeleChat Community

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

31,768 INPUT + 1,000 OUTPUT · 3.22P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT31.8KTEXT

TRANSFORMER STACK64 DENSE LAYERSRMSNorm · GQA · SwiGLU in every block

Pre-Norm Residual

OUTPUT1.0KTOKENS

wraps every repeated block

DENSE GQA · TELECHAT YARN位置插值把原生 8K 扩展到 32K，而不改变 attention 图。

Dense SwiGLU没有 MoE 路由，所有 FFN 权重对每个 token 都激活。

DEPTH64dense

HIDDEN6,144trunk

HEADS48 / 8Q / KV

FFN24,576SwiGLU

CONTEXT32KYaRN

PARAMS36.24Bheader