Skip to main content

Tencent Zixiao 2.0

Product Overview​

Tencent Zixiao 2.0 is Tencent's sixth-generation custom AI chip, officially unveiled by Tencent Cloud on July 20, 2026, the third day of WAIC 2026 in Shanghai. It is the second-generation product of Tencent's in-house "Zixiao" chip line. Compared with the first-generation Zixiao released in 2021, the most important positioning change is expanding from a pure inference chip to a unified training-and-inference design — Tencent officially describes Zixiao 2.0 as targeting both large-model training and inference.

Performance figures (official Tencent Cloud data): Zixiao 2.0 delivers 4x the training performance and 3x the inference throughput of the previous generation. In deployment, Tencent's internal businesses were the first to adopt Zixiao 2.0, achieving a 50% performance improvement in Hunyuan large-model training.

Development model: Zixiao 2.0 is a joint development effort between Tencent and Enflame Technology (燧原科技). Tencent's Penglai Laboratory leads chip R&D, with Enflame (in which Tencent holds roughly 20.258%, making it Tencent's largest external shareholder) deeply involved. This mirrors the first-generation Zixiao model — Tencent combining capital ties with joint development to supplement its in-house capabilities, while Enflame also supplies Tencent's training compute.

Deployment plans: Tencent Cloud announced that custom Zixiao chips are planned to reach 30% of new custom-chip additions in 2026, and that total custom-chip deployment is planned to expand to a 100,000-accelerator scale by the end of 2026. Over the same period, Tencent Cloud's HCC high-performance computing cluster completed an upgrade now supporting 100,000-accelerator LLM training.

⚠️ Critical disclosure boundary (this site's core principle)

As with the 2021 first-generation Zixiao, Tencent has to date published none of Zixiao 2.0's hard specifications — process node, memory capacity, memory type, memory bandwidth, FP8/FP4/FP16/FP32/INT8 compute, TDP, and chip interconnect are all undisclosed. Tencent has only published three relative multiples (4x training, 3x inference throughput, 50% Hunyuan training gain) plus deployment scale plans.

Circulating claims such as "Zixiao C200" or "48GB memory per accelerator" originate from third-party media and independent blogs. Tencent has never confirmed them, and they have no counterpart in official release information. This page does not treat them as fact and excludes them from the specification table.

Every field requiring an absolute number is marked "not disclosed". Compared with inserting unsourceable numbers, honestly flagging undisclosed information better reflects the real state of the data — consistent with how this site treats the first-generation Zixiao card.

Key Specifications​

ItemSpecification
ArchitectureTencent in-house AI chip architecture (unified training and inference; microarchitecture details undisclosed)
ProcessNot disclosed
MemoryNot disclosed
Memory TypeNot disclosed
Memory BandwidthNot disclosed
FP8Not disclosed
FP4Not disclosed
FP16Not disclosed
FP32Not disclosed
INT8Not disclosed
TDPNot disclosed
InterconnectNot disclosed
Release / ProductionReleased 2026-07-20 (WAIC 2026, Shanghai); Tencent plans to expand deployment to a 100,000-accelerator scale by end-2026

⚠️ The only officially published performance data are relative multiples:

Official figureValueNote
Training performance gain4xversus previous Zixiao generation
Inference throughput gain3xversus previous Zixiao generation
Hunyuan training gain50%realized business-side benefit
2026 new custom-chip share30%Tencent Cloud deployment plan
End-2026 deployment scale100,000-accelerator classTencent Cloud deployment plan

Important: "4x training performance" does not mean "per-accelerator compute is 4x the previous generation" — the comparison baseline, precision basis (FP8 / FP16 / INT8), and measurement method (single accelerator vs. cluster) are all undisclosed. Without this information any absolute compute derivation would be unreliable.

Key Features​

  • Unified training and inference: expanding from the first generation's pure inference chip to cover both large-model training and inference is Zixiao 2.0's most important positioning shift
  • Official performance claims: 4x training, 3x inference throughput (both relative to the previous generation)
  • Already deployed in Hunyuan: the 50% training gain in Tencent's Hunyuan model represents a realized, production business benefit
  • Tencent + Enflame joint development: continuing the collaboration begun with Tencent's Enflame investment in 2018 (Tencent holds roughly 20.258%, its largest external shareholder)
  • 100,000-accelerator deployment plan: Tencent targets 30% custom Zixiao share of new additions in 2026 and 100,000-accelerator custom-chip deployment by end-2026
  • Ecosystem collaboration: at WAIC 2026, Tencent Cloud jointly showcased large-model appliances and intelligent computing clusters with Enflame, Iluvatar CoreX (天数智芯), and MetaX (沐曦)
  • Strategic significance: reducing reliance on NVIDIA GPUs, forming the core of Tencent's parallel "in-house R&D + investment + procurement" compute strategy
  • Limitations: no hard specifications disclosed, preventing compute-level comparison with any comparable chip; not sold as standalone silicon

Use Cases​

  • ✅ Large-model training and inference inside Tencent (already deployed for Hunyuan)
  • ✅ Tencent Cloud public AI compute supply (accessible via cloud service; silicon not sold directly)
  • ✅ Domestic compute substitution scenarios (enterprise and government customers with high NVIDIA GPU dependency)
  • ✅ Intelligent computing cluster buildouts (joint solutions with Enflame, Iluvatar CoreX, and MetaX, complementing the 100,000-accelerator HCC cluster)
  • ❌ Specification-based procurement comparisons (all hard metrics undisclosed)
  • ❌ Direct purchase of chips or boards (Tencent does not retail; access is via Tencent Cloud only)
  • ❌ Drop-in inference replacement (3x inference throughput over its own previous generation, but no absolute performance data)

Comparison​

Comparison note: Zixiao 2.0 and the first-generation Zixiao can be compared directly on official claims (both are Tencent's own relative multiples). However, Zixiao 2.0 cannot be placed into a compute comparison table alongside chips that do publish hard specifications, such as Hanguang 800 or MLU 370 — this is the most significant difference in public information between Tencent's in-house chips and domestic peers.

Data Sources​