China’s AI labs are not one bloc — Alibaba, DeepSeek, Moonshot and Ant are betting differently
Alekseener33 · reddit · 2026-07-27
A Reddit post argues that the Chinese AI labs people often lump together are actually making very different bets.
The author splits them into four main strategies:
- Alibaba/Qwen: distribution and ubiquitous support across sizes, quantizations, and runtimes.
- DeepSeek: architecture-first, publishing the paper and weights together.
- Moonshot: a longer-horizon approach that can look odd for a release cycle if it pays off later.
- Ant Ling: serving cost, with Ling-3.0-flash described as a 124B MoE model with about 5.1B active parameters per token, KDA+MLA hybrid attention, and 262K context, optimized for long agent loops at low cost rather than leaderboard wins.
The post also notes that Ant announced before opening weights, which helps infra teams but reduces goodwill among people who want to run the model locally. The author asks whether knowing the specific lab changes how others interpret these releases.
More from Companies & People
- OpenAI co-hosted GPT-6 hackathons in SF and NYC with Cerebral Valley — OpenAIDevs · 2026-09-23
- Meta's Alexandr Wang reveals muse has been in the works since at least Sept 2025 — adrianscottcom · 2026-09-23
- OpenAI shares GPT-6 hackathon tale: builder used real star maps with Astra to find way home — OpenAIDevs · 2026-09-23
- Instacart lands on Meta's Muse: say "Taco Tuesday" and get an auto-built grocery cart — alexandr_wang · 2026-09-23
- Databricks ships GPT-6 Sol/Luna and Claude Opus 5.5 with Unity Gateway model governance — matei_zaharia · 2026-09-23
- Fields Medalist Martin Hairer Explains Why AGMAI Is Truly Independent of OpenAI — AlexKontorovich · 2026-09-23