Tencent Stacks Model from 295B to 770B in 6 Weeks — China's Open-Source Sprint

What this is

This week Tencent pushed its open-weights model from 295B to 770B parameters — a 2.6x jump in six weeks. The Hy4 preview is live on Hugging Face: 770B total parameters, 49B active (MoE, or Mixture of Experts — only a subset of parameters fires on each inference pass), 1 million token context window, text-only input, with model files totaling 1.56TB. Compared with July's Hy3 (295B total, 21B active, 256K context), scale and context both reach a new tier.

Industry view

We see two camps forming. Optimists argue: Chinese majors reaching this scale in six weeks and shipping open weights is a substantive shock to the global mid-tier AI application market — enterprises can self-host a 1M-context model cheaply for contracts, research reports, and long documents; with 49B active parameters under MoE, inference cost is not prohibitive.

Skepticism is sharp, and we think the criticisms land. First, "770B" is a marketing number — the figure that actually drives inference cost is 49B active, which is not in the same league as a true 770B dense model. Second, open weights ≠ open source: no training data, no fine-tuning recipes, no alignment methodology — enterprises have limited room to customize. Third, text-only input is a clear shortcoming versus multimodal models shipping in the same window. Fourth, the team has published no benchmark results — the only public output so far is a single "pelican riding a bicycle" SVG generation, which tells us very little.

Impact on regular people

For enterprise IT: another open-weights option to evaluate, especially for self-hosted deployments and document-heavy workloads — but be cautious, building your own inference cluster may not pencil out for SMBs.

For working professionals: long context plus strong text reasoning delivers real value for tasks that need entire research reports, contracts, or books fed in at once; for most users, calling the API beats self-hosting.

For consumers: continued scaling of open-source LLMs keeps pushing overall inference cost down — there's still room for AI app subscription and API prices to fall.