Tencent¶

Development line: organization:tencent · thread ai-research-releases
Last event: 2025-12-29 · 1 dated since 2025-12-29 · Researched: 2026-09-04 · confidence: medium

What it is¶

Tencent sells consumer platforms, cloud, marketing, payments and digital services to consumers and enterprises.

• Consumer and enterprise product lines are separate.
• TencentARC’s DSR Suite is an Apache-2.0, Qwen2.5-VL-7B-derived VLM package for reasoning about object geometry and relationships over time in video.
• It includes a 50K-QA training set, a 1,484-QA benchmark, a 9B BF16 model and a Geometry Selection Module.

Development line¶

• 2025-12-29 — Tencent ARC published the DSR Suite model and code repositories. We get code and weights for dynamic spatial reasoning in video, with DSR-Train, DSR-Bench and GSM on a Qwen2.5-VL-7B base.

What changed¶

Tencent development line:

• 2023-05-02 — Tencent Cloud reported a low-sample digital-human production platform as a self-service avatar workflow. The report said it could use a three-minute speaking video and 100 voice sentences. This covers a Tencent Cloud product, not DSR Suite.
• 2025-12-29 — TencentARC released DSR Suite as open research. It provides code and weights for dynamic spatial reasoning in video, using DSR-Train, DSR-Bench and GSM on a Qwen2.5-VL-7B base.
• 2025-12-25 to 2026-02-21 — The repository dates the paper and code to 2025-12-25, and lists CVPR 2026 acceptance on 2026-02-21. That changes academic standing, not availability as a managed Tencent product.

How to use this¶

As of 2025-12-29, begin with the published DSR Suite model and source repositories, then verify capabilities and licensing before adoption.

1. Choose the named Tencent product line first; Tencent itself is not one developer endpoint, and its public catalog separates consumer services, Tencent Cloud and marketing products. — https://www.tencent.com/what-we-create/
2. For DSR Suite local inference, follow the model card’s Transformers loader for AutoProcessor and Qwen2_5_VLForConditionalGeneration_Spatial, loading TencentARC/DSR_Suite-Model with device_map="auto". — https://huggingface.co/TencentARC/DSR_Suite-Model
3. For a reproducible check, download DSR-Bench as benchmark.parquet, set local video and parquet paths in VLMEvalKit, then run its Spatial-Reasoning task. — https://raw.githubusercontent.com/TencentARC/DSR_Suite/main/README.md
4. For fine-tuning, create the separate Python 3.11 model environment, install model requirements, obtain video/caption data and optional DSR-Train, convert QA pairs, set the QA, video and Pi3 checkpoint paths, then run train.sh. — https://raw.githubusercontent.com/TencentARC/DSR_Suite/main/model/README.md

Best practices¶

• Run data generation and model training in separate environments because package requirements differ. — https://raw.githubusercontent.com/TencentARC/DSR_Suite/main/README.md
• Plan for local hosting of 9B BF16 weights rather than assuming a hosted endpoint; the checked model card lists no inference provider. — https://huggingface.co/TencentARC/DSR_Suite-Model
• Evaluate with DSR-Bench only after obtaining the referenced videos and setting both local paths; the benchmark parquet does not supply video files directly. — https://raw.githubusercontent.com/TencentARC/DSR_Suite/main/README.md
• For custom QA generation, remove static scenes before generation and configure API access: the curation scripts require your own endpoint and token. — https://raw.githubusercontent.com/TencentARC/DSR_Suite/main/data/README.md

Superseded by this¶

• No verified product supersession: the 2025-12-29 DSR Suite release is not documented as a successor to the 2023-05-02 Tencent Cloud low-sample digital-human platform.
• 2026-02-21: CVPR 2026 acceptance supersedes describing the 2025 paper only as a pre-acceptance release; it does not establish a new managed service.

Still unknown¶

• The 2023 Tencent Cloud digital-human report and the 2025 TencentARC DSR Suite release are different products and teams; no checked source establishes a shared API, product lineage or replacement relationship.
• No current first-party public page was verified for access, price, regional availability or support status of the 2023 digital-human platform.
• The checked DSR Suite sources do not state a tested Transformers version, GPU-memory requirement or Tencent Cloud deployment path.

Sources¶

Agent brief¶

• Subject: organization:tencent, thread ai-research-releases, 1 dated events 2025-12-29 → 2025-12-29.
• Practical note: As of 2025-12-29, practitioners evaluating Tencent ARC work should begin with the published DSR Suite model and source repositories, then independently verify its capabilities and licensing before adoption.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.