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
| Parameter | Value |
|---|
| Positioning | Dedicated to large-scale AI inference |
| Architecture | In-house architecture (specific codename not disclosed) |
| Process | Not disclosed |
| FP16 / BF16 | Not disclosed |
| INT8 / INT4 | Not disclosed |
| Memory Capacity | Not disclosed |
| Memory Type | Not disclosed |
| Bandwidth | Not disclosed |
| TDP | Not disclosed |
| Interconnect | Tianchi supernode ecosystem |
| Announced | November 13, 2025 (Baidu World) |
| Launch | Planned for early 2026 |
| Production Status | Commercial 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
| Parameter | Details |
|---|
| Company | Kunlunxin Technology (Beijing) Co., Ltd. |
| Parent Company | Baidu (57.67% stake) |
| M100 Announcement | Baidu World, November 13, 2025 |
| IPO Status | Started STAR Market IPO tutoring in May 2026 |
| Deployment Scale | Tens of thousands of cards deployed across the Kunlunxin lineup |
| Core Scenario | Inference 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
| Dimension | M100 (Inference) | P800 (Training + Inference) |
|---|
| Positioning | Inference-dedicated | General training + inference |
| MoE Optimization | Native optimization | Supported |
| PD Disaggregation | Supported (+95% performance) | Basic support |
| Single-Server Deployment | Cloud inference services | 8-card 671B in a single server |
| Launch | Early 2026 | 2024-03 |
| Supernode | Tianchi 256/512 | Tianchi 256/512 |
| Spec Disclosure | To be announced | Published |
Key Timeline
| Date | Event |
|---|
| 2024-03 | P800 launched |
| 2025-04 | Tianchi supernodes enabled on Baige 5.0 |
| 2025-11-13 | M100/M300 announced (Baidu World) |
| 2026-01 | M100 entered its commercial volume phase |
| H1 2026 | M100 official mass production and delivery |
| Early 2027 | M300 launch (trillion-parameter-scale training) |