GLM-5.3 just went open source a week ago, and Tencent's 770B Hy4 has already caught up
GLM-5.3 just went open-source a week ago, and Tencent's 770B Hy4 is already catching up. The domestic large model race has been intense this week. On August 27, Zhipu open-sourced GLM-5.3-Flash, with benchmark scores matching Opus 4.8 and a price of 0.4 yuan per million tokens—I even wrote a dedicated piece about it yesterday. Then the very next day, August 28, Tencent Hunyuan dropped its new flagship model.
💡 What You Will Learn
GLM-5.3 just went open-source a week ago, and Tencent's 770B Hy4 is already catching up. The domestic large model race has been intense this week. On August 27, Zhipu open-sourced GLM-5.3-Flash, with
📜 Table of Contents
GLM-5.3 Just Went Open-Source a Week Ago, and Tencent's 770B Hy4 Is Already Catching Up
Domestic large models have been going at it pretty hard this week.
On August 27, Zhipu open-sourced GLM-5.3-Flash, with benchmark scores matching Opus 4.8 and a price of 0.4 yuan per million tokens—I even wrote a dedicated piece about it yesterday. Then the very next day, August 28, Tencent Hunyuan dropped its new flagship model Hy4: 770B parameters, 1M context, also open-source, and priced just as aggressively.
But one detail caught my eye: this time, Tencent didn't lead with benchmark leaderboards—they led with a "blind test."
Tencent Skips the Benchmarks This Time, Shares a Blind Test Instead
According to Tencent Hunyuan's official blog, they gathered 163 internal experts and designed 203 real-world engineering tasks, putting Hy4, GLM 5.3, and Kimi K3 head-to-head across coding, office work, and scientific tasks. The experts didn't know which model produced which answer, and scored on a 4-point scale.
The result: Hy4 averaged 2.99, Kimi K3 came in at 2.94, and GLM 5.3 scored 2.92.
It's a subtle way to win. Against GLM 5.3 specifically, Hy4's win rate was 46.8%, ties at 12.8%, and losses at 40.4%. Calling it "slightly better" is accurate; calling it a "crushing victory" would be pure hype.
Why Big Tech Is Falling Out of Love with Benchmark Leaderboards (The Evaluation Wars)
Public benchmarks (those standardized exam-style tests) have developed an increasingly obvious flaw over the past couple of years: the questions are public, so models can easily "game" them during training by seeing similar problems. Scores keep climbing, but real-world performance doesn't necessarily follow. It's the same logic as students cramming for exams—ace every mock test, and still choke on the actual job.
Tencent took a different approach this time: skip the standardized tests entirely and pull tasks straight from internal, real business workflows—writing code, building spreadsheets, running scientific computations—then let experts who do this stuff every day score the outputs blind. Whether a model is actually good? The people who use it for real work get the final say.
This evaluation method is closer to how you and I actually use models than any leaderboard, but it has its own limitations: the test was organized by Tencent itself, the experts are Tencent employees, the tasks were chosen by Tencent, and the GLM and Kimi models they compared against are the current live versions. It shows Hy4 performs well in Tencent's productivity scenarios—but conclusions like "best in China" will need more independent testing to back up.
Hard Specs and Pricing (Free for a Limited Time)
Quick rundown of the specs. Hy4 uses a MoE architecture (Mixture of Experts—think of it as a model packed with multiple specialized "brains," where only the relevant ones get activated per task), with 770B total parameters, 49B activated per inference, and a 1M context window—one million tokens, roughly enough to fit an entire novel in one go. The previous generation, Hy3, had 295B total parameters, 21B activated, and a 256K context—so Hy4 roughly doubles or more every spec across the board.
API pricing per million tokens: 0.3 yuan for cache-hit input, 6 yuan for standard input, and 18 yuan for output. Model weights are open-sourced across four platforms including GitHub and HuggingFace.
There's also a genuinely useful perk tied to the release: Tencent's AI office assistant WorkBuddy has already integrated Hy4, and it's free from August 28 to September 10—no token costs during that window, so you can try it hands-on right away. The free period for the previous-gen Hy3 has also been extended to September 30. Apps like Yuanbao and ima offer free access too.
Tencent was also upfront about a caveat: this is a preview build, and on complex tasks it can "overthink" and repeatedly second-guess itself. The full release is still being iterated on.
The Selection Logic Has Shifted: From Scores to Real Work
Put this week's two events side by side and the trend is clear: Zhipu is showing public benchmarks matching overseas flagships; Tencent is showing internal blind tests slightly edging out domestic rivals. Two different plays—one proving "we're not worse than the international guys," the other proving "we're better at real work."
For everyday users like us, the real good news is that domestic open-source models are competing this fiercely—the strong keep getting stronger, prices keep dropping, and options keep multiplying. Take leaderboards with a grain of salt. Whether a model can actually finish that chunk of code or that spreadsheet for you—that's what really counts.
It's 2026 now—do you use large models in your daily work? Compared to GPT and Claude, how big do you think the gap still is for domestic models? Drop your thoughts in the comments.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
