Skip to main content

Kunlunxin M100 (2026)

Product Overview​

The Kunlunxin M100 is a new-generation AI inference chip unveiled by Kunlunxin Technology at Baidu World on November 13, 2025, designed and optimized for large-scale AI inference scenarios, especially inference on MoE (Mixture of Experts) architecture models. It is scheduled to launch in early 2026 and entered its commercial volume phase in January 2026.

Note: M100's detailed hardware specs (compute, memory, power, etc.) were not disclosed at launch; the information below is compiled from official announcements and industry reports.

M Series Positioning:

  • Kunlunxin M100 (early 2026): large-scale AI inference — this page
  • Kunlunxin M300 (early 2027): ultra-large-scale multimodal LLM training and inference
  • Kunlunxin P800 (2024): general training + inference accelerator card — existing page
  • Kunlunxin N Series (2029): next-generation architecture

Core Specifications​

ParameterValue
PositioningDedicated to large-scale AI inference
ArchitectureIn-house architecture (specific codename not disclosed)
ProcessNot disclosed
FP16 / BF16Not disclosed
INT8 / INT4Not disclosed
Memory CapacityNot disclosed
Memory TypeNot disclosed
BandwidthNot disclosed
TDPNot disclosed
InterconnectTianchi supernode ecosystem
AnnouncedNovember 13, 2025 (Baidu World)
LaunchPlanned for early 2026
Production StatusCommercial volume phase since January 2026

Key Features​

  • MoE inference optimization: hardware-level optimization for the sparse activation characteristics of MoE, significantly boosting MoE model inference performance
  • PD-disaggregated inference: supports Prefill-Decode disaggregated deployment, raising single-card performance by 95%
  • Single-instance performance: up to an 8x improvement when combined with inference optimizations
  • Tianchi supernodes: works with the Tianchi 256 / Tianchi 512 supernodes to build thousand-card inference clusters
  • China Mobile win: first place in share for the CUDA-ecosystem lot of the inference-type centralized procurement

Vendor Information​

ParameterDetails
CompanyKunlunxin Technology (Beijing) Co., Ltd.
Parent CompanyBaidu (57.67% stake)
M100 AnnouncementBaidu World, November 13, 2025
IPO StatusStarted STAR Market IPO tutoring in May 2026
Deployment ScaleTens of thousands of cards deployed across the Kunlunxin lineup
Core ScenarioInference service foundation of Baidu AI Cloud

Use Cases​

  • ✅ Large-scale AI inference (LLM online services)
  • ✅ MoE model inference (hardware optimization for sparse activation)
  • ✅ PD-disaggregated deployment (independent optimization of Prefill + Decode)
  • ✅ Baidu Cloud inference services (inference for Qwen, ERNIE, and other models)
  • ✅ Domestic inference clusters
  • ❌ AI training (positioned as inference-dedicated; use the P800/M300 for training)
  • ❌ Specs to be confirmed (watch the official 2026 product launch for detailed parameters)

Positioning Comparison with the P800​

DimensionM100 (Inference)P800 (Training + Inference)
PositioningInference-dedicatedGeneral training + inference
MoE OptimizationNative optimizationSupported
PD DisaggregationSupported (+95% performance)Basic support
Single-Server DeploymentCloud inference services8-card 671B in a single server
LaunchEarly 20262024-03
SupernodeTianchi 256/512Tianchi 256/512
Spec DisclosureTo be announcedPublished

Key Timeline​

DateEvent
2024-03P800 launched
2025-04Tianchi supernodes enabled on Baige 5.0
2025-11-13M100/M300 announced (Baidu World)
2026-01M100 entered its commercial volume phase
H1 2026M100 official mass production and delivery
Early 2027M300 launch (trillion-parameter-scale training)