Hunyuan-A13B Technical Report

TL;DR

Tencent's technical report for Hunyuan-A13B, an open-source MoE with 80B total parameters activating 13B at inference, pretrained on a rigorously filtered 20T-token corpus with enhanced STEM data curation, and carrying a dual-mode Chain-of-Thought framework — fast thinking for routine queries, slow thinking for complex multi-step problems (source).

Authors & Org

Not stated in the snapshot — no author list. arxiv.org answers EGRESS_BLOCKED from this run's sandbox.

The report is Tencent's by its subject. That page records that Hunyuan's prior publications were credited to the "Tencent Hy Team" and to Tencent Hunyuan and name no individual, which is consistent with what is readable here.

Method

Stated architecture and training:

Element	Stated value
Architecture	Mixture-of-Experts
Total parameters	80B
Active at inference	13B
Pretraining corpus	20T tokens, "rigorously filtered", enhanced STEM data curation
Post-training	high-quality supervised fine-tuning + large-scale reinforcement learning
Reasoning	dual-mode Chain-of-Thought: fast thinking / slow thinking, adapting reasoning depth to task complexity
The stated design goal is **"balancing model capability, computational efficiency, and
deployment cost"**, with high inference throughput making it **"suitable for
latency-sensitive applications"**.

The 80B/13B ratio is the specification claim: a 6.2× gap between total and active parameters, which is where the cost argument comes from.

Not stated in the snapshot: expert count, expert routing, context window, layer count, tokenizer, the filtering criteria, the RL method, or what triggers the switch between fast and slow thinking. Developer aside, nothing here is enough for a model page — see below.

Results

No figure of any kind appears in anything read. Performance is described qualitatively: "competitive performance across mathematics, science, programming, general language understanding, and agent tasks, often approaching that of much larger models".

Five domains named, zero scores. The comparison — "approaching much larger models" — names no model and no benchmark.

This repo's own captured leaderboard does not list it. The Artificial Analysis table read 2026-09-27 carries exactly one Tencent row, Hy3 at Intelligence Index 25, 256k context, $0.07 Cost per Task USD, and no Hunyuan-A13B (source). So the report's claims cannot be checked against the one independent instrument this repo captures.

Significance

The report's date and the model's date do not agree, and that is the substance of this page.

The HuggingFace Daily Papers snapshot dates this entry 2026-09-23 (source). Every search pass run on 2026-09-27 states that Hunyuan-A13B's weights were open-sourced on 2025-06-27 — pretrained and instruct weights together — giving a 256K context window, dual reasoning modes, and a custom Tencent licence whose territory excludes the European Union, the United Kingdom and South Korea.

If both are right, this is a technical report published roughly fifteen months after the weights, which is unusual and not impossible. Nothing read states it, so it is recorded in ## Open Questions as an unresolved conflict rather than smoothed over.

No model page was created, and the date is why. This wiki's schema requires a Released value, and the two candidates are fifteen months apart: the arXiv date the snapshot gives, and a 2025 date that appears only in search summaries with no first-party read behind it (huggingface.co and github.com were not reached). A model page asserting either would be this wiki publishing a release date it cannot support, on a page whose whole purpose is to be the citable record of one. The material is recorded here and on Tencent instead, per the one-off-mention rule — and it is the right call to revisit the moment a first-party read is possible.

For Tencent this is the first entry since 2026-09-01, and the page's own standing observation was that the lab is held here largely as leaderboard rows: Hy4 preview from the 770B release that created the page, and Hy3 known only as an Artificial Analysis line. That has not changed — a 20T-token, 80B-parameter open MoE with no published number is still not a model this wiki can place against anything.

Against Ling-3.0-tiny, captured the same run: two Chinese open-weight MoE releases reaching this wiki on the same day, both with no citable benchmark, one at 7.9B/1.3B and one at 80B/13B. Ling 3.0 Tiny at least has a captured leaderboard row with the publisher's column heading above it. This has nothing.

Open Questions

• When was Hunyuan-A13B released? The snapshot dates the report 2026-09-23; every search pass dates the weights 2025-06-27. Unresolved, and it blocks a model page.
• What are the numbers? Five domains claimed, no score in any of them, no named benchmark, no named comparator.
• What is the licence? Search passes give a custom Tencent licence excluding the EU, UK and South Korea — a materially restrictive term for something called open-source, and not read first-party. It bears directly on Open-Weights Policy Fight if confirmed.
• What is the context window? 256K per search passes; absent from the snapshot.
• What switches fast thinking to slow? The dual-mode CoT framework is the report's distinguishing feature and its trigger is unstated.
• Why is it not on the leaderboard this repo captures? An 80B open MoE absent from a 269-row table is itself a signal about uptake.

Cite

arXiv 2609.27284 — Hunyuan-A13B Technical Report, 2026-09-23. HuggingFace Daily Papers, 2026-09-27, 10 upvotes — a popularity signal from that community and not a quality or importance ranking (source).

Not read first-party: the GitHub repository (GitHub) and the Hugging Face model card (Hugging Face) — huggingface.co answers connect_rejected under standing policy and github.com was not reached this run.