Skip to main content

Moore Threads Huashan (Fifth-Generation "Huagang" Architecture · Unified AI Training and Inference)

Product Overview​

Huashan is Moore Threads' high-performance unified AI training-and-inference chip built on the fifth-generation "Huagang" architecture, with mass production planned within 2026.

Huashan succeeds the current cloud mainstay MTT S5000 (fourth-generation MUSA architecture). At Moore Threads' 2026 interim results briefing, the company disclosed that its R&D focus is "the Huashan and Lushan chips under the new-generation 'Huagang' architecture and related applications" — Huashan handles cloud AI training and inference while Lushan handles graphics rendering, both sharing the Huagang architecture foundation.

Huashan's industrial significance lies in carrying forward the existing ecosystem: the MTT S5000 has already supported the 10,000-card-scale "Kuae (KUAE)" intelligent computing cluster in training the MoE-236B model from scratch, as well as full-stack native training of the world's first 5D world model, EvoPhys-World. Building on this, Huashan must further improve per-card compute and cluster interconnect efficiency to sustain Moore Threads' technical goal of a 100,000-card-scale intelligent computing cluster.

Core Specifications​

ParameterValue
ArchitectureFifth-generation "Huagang" architecture
Process NodeNot disclosed
PositioningUnified cloud large-model training + inference
FP16/BF16 ComputeNot disclosed
FP8 ComputeNot disclosed
Memory CapacityNot disclosed
Memory TypeNot disclosed
Memory BandwidthNot disclosed
TDPNot disclosed
InterconnectMoore Threads proprietary cluster interconnect (supporting 10,000-card-scale clusters)
Software StackMUSA architecture (fifth generation)
LaunchMass production within 2026 (in development)

⚠️ Specification Notes: Huashan is currently in development and Moore Threads has not published complete specifications. All compute, memory, bandwidth, and power fields are not disclosed; do not substitute speculative values for official specifications.

Huashan's Ecosystem Inheritance: From the MTT S5000 to 10,000-Card Clusters​

Moore Threads has already run the complete "domestic chip training domestic model" pipeline with its previous-generation product, and this engineering accumulation will migrate directly to Huashan:

CapabilityAchieved on MTT S5000Huashan's Goal
Model training10,000-card-scale "Kuae" cluster completed from-scratch training of MoE-236BLarger-scale cluster training
World modelFull-stack native training of the world's first 5D world model EvoPhys-WorldContinue multimodal capability
InferenceInference of mainstream models such as GLM / DeepSeek, PD heterogeneous disaggregation challenging 1M ultra-long contextImprove inference efficiency and cost
Cluster scale10,000-card-scale intelligent computing cluster100,000-card-scale technical foundation
Software stackMUSA + Kuae cluster software stackFifth-generation MUSA

Relationship with the Current Mainstay MTT S5000​

ParameterMTT S5000Huashan
Architecture generationFourth-generation MUSAFifth-generation Huagang
StatusIn mass production (stable supply)Mass production within the year (in development)
PositioningUnified cloud training and inferenceUnified cloud training and inference (successor)
SpecificationsPublicNot disclosed
Purchase adviceAvailable nowEvaluate after official release

Key Insight: Huashan follows the same logic as the MTIA 400 for Meta and the TPU 8t for Google — self-developed chips replace the current generation with the next, so buyers must judge timing between "the S5000 they can buy today" and "the Huashan with unknown performance". In terms of immediate deployability, the MTT S5000 remains the most mature choice among domestic full-function GPUs.

Vendor Information​

ParameterValue
CompanyMoore Threads Intelligent Technology (Beijing) Co., Ltd.
HeadquartersBeijing
Founded2020-10
ListingSTAR Market (688795)
Company PositioningFull-function GPU (graphics + AI compute)
Cluster SolutionKuae (KUAE) intelligent computing cluster
Software StackMUSA architecture

Use Cases​

  • ✅ Cloud large-model training (inheriting the 10,000-card-scale cluster solution)
  • ✅ Large-model inference (including long context and PD disaggregation architecture)
  • ✅ Domestic substitution in intelligent computing centers
  • ❌ Professional graphics rendering (left to Lushan)
  • ❌ No product currently on sale (in development)

References​