VC
Value Add VC
⚡HomePulse⚡Helpful Apps📝Blog
← Value Add PulseAI295B model, 21B active, beats larger rival on most benchmarks

Tencent's Hy3 Beats GLM-5.2 at Half the Size

Tencent released Hy3, an Apache-licensed 295B open model with 21B active parameters, matching or beating GLM-5.2's roughly 744B model on most benchmarks except coding, intensifying China's open-weight model race.

295B (21B active)
Hy3 Size
~744B (40B active)
GLM-5.2 Size
90.4
GPQA Diamond
Apache 2.0
License
TC
Trace Cohen
Early-stage VC & angel · Founder, New York Venture Partners
July 6, 2026
1 min read
ShareXLinkedInEmail
THE RUNDOWN
1

Hy3 is a 295B mixture-of-experts model with 21B active parameters and 256K context, released July 6 under a permissive Apache 2.0 license, versus GLM-5.2's roughly 744B total, 40B active parameters

2

Hy3 posts 90.4 on GPQA Diamond and 72.0 on USAMO 2026, and leads GLM-5.1 in a blind 270-expert evaluation across real-world workflows, 2.67 versus 2.51 out of 4

3

GLM-5.2 still leads on coding benchmarks specifically, while Hy3 leads the open field on agentic search and tool orchestration, posting 84.2 on BrowseComp and 91.0 on DeepSearchQA

4

The efficiency gap -- comparable or better performance at less than half the active-parameter footprint -- directly pressures the cost structure closed, API-only frontier labs can sustain

TC
The VC Read · Trace's TakeTrace Cohen

Half the active parameters for comparable performance is the number that should worry every closed-model API business, not the benchmark table itself -- efficiency gains like this compound directly into inference cost, and inference cost is the actual battleground for enterprise AI spend in 2026. Founders building anywhere near the model layer should assume the open-weight efficiency curve keeps bending down faster than closed-lab pricing can follow, and plan accordingly.

Tencent released Hy3 on July 6, an open-weight, Apache 2.0-licensed model built as a 295 billion parameter mixture-of-experts architecture with only 21 billion parameters active at inference time and a 256K context window. The efficiency framing is the point: Hy3 matches or beats Zhipu's GLM-5.2, a roughly 744 billion parameter model with 40 billion active parameters, on most benchmarks while running at less than half the active-parameter footprint.

On STEM and reasoning tasks, Hy3 reports 90.4 on GPQA Diamond, 72.0 on USAMO 2026, and 90.0 on IMOAnswerBench. In a blind evaluation Tencent ran with 270 experts across disciplines working real-world workflows, collecting 312 valid comparisons, Hy3 scored 2.67 out of 4 against GLM-5.1's 2.51, with its clearest advantages in frontend development, CI/CD and data-and-storage tasks.

The one category where GLM-5.2 still clearly leads is coding accuracy across the full benchmark suite -- Hy3 trades some coding performance for its dramatically smaller active footprint. But on agentic search and tool orchestration, Hy3 leads the entire open field, posting 84.2 on BrowseComp and 91.0 on DeepSearchQA, with a leading 79.1 on the public MCP-Atlas tool-orchestration benchmark.

“On STEM and reasoning tasks, Hy3 reports 90.4 on GPQA Diamond, 72.0 on USAMO 2026, and 90.0 on IMOAnswerBench.”

The release keeps China's open-weight race compounding: Tencent, Zhipu's GLM series, Alibaba's Qwen family and DeepSeek have all shipped rapid iterations through 2025 and 2026, a category where Western frontier labs like OpenAI and Anthropic remain comparatively less active, generally preferring closed, API-only deployment for their most capable models.

For infrastructure and applications investors, the efficiency angle matters more than the raw benchmark scores: a model matching a much larger rival's performance at under half the active-parameter cost directly lowers inference costs for anyone deploying it, and every such release compounds pricing pressure on closed frontier-lab APIs regardless of where the end customer sits.

What to watch: whether Hy3 closes the remaining coding gap with GLM-5.2 in a near-term follow-up release, and whether Western closed-model labs respond with their own efficiency-focused releases rather than continuing to compete purely on raw capability.

ShareXLinkedInEmail
More onTencent →

Originally reported by VentureBeat. Analysis and editorial commentary by Value Add Pulse.

← Back to Pulse

THE WIRE in your inbox

Tech, startup & VC news with Trace's take. Free, no spam.

Read Next

AI$19B lease, 20-year term

Anthropic Signs $19B Data-Center Lease, Files for IPO

Anthropic signed a 20-year, $19 billion data-center lease with TeraWulf in Kentucky, backed by credit it doesn't yet have, the same month it confidentially filed for an IPO expected to price this fall.

AINew interpretability finding, no product change

Anthropic Finds a 'Workspace' Inside Claude's Mind

Anthropic's new J-lens interpretability technique found a privileged internal 'workspace' in Claude that mirrors Global Workspace Theory, a leading model of human consciousness, without claiming the model is actually conscious.

AICorrection: human-directed, not fully autonomous

'First AI-Run Ransomware Attack' Still Needed a Human

New reporting on the widely cited 'first autonomous AI ransomware attack' finds a human operator still directed key decision points, complicating the fully-autonomous framing that spread when the incident first surfaced.

@Trace_Cohen·t@nyvp.com