VC
Value Add VC
⚡HomePulse⚡Helpful Apps📝Blog
← Value Add PulseAIWins most benchmarks at ~50% size

Tencent's Hy3 Takes on GLM-5.2 at Half the Size

Tencent's Apache-licensed Hy3 model wins most benchmarks against GLM-5.2 at roughly half the parameter count, though it still trails on coding tasks, intensifying China's open-weight model race.

Apache 2.0
License
~50%
Size vs. GLM-5.2
Coding tasks
Weak Spot
TC
Trace Cohen
Early-stage VC & angel · Founder, New York Venture Partners
July 6, 2026
1 min read
ShareXLinkedInEmail
THE RUNDOWN
1

Tencent released Hy3 under an Apache license, positioning it as a directly comparable, freely usable alternative to Zhipu AI's GLM-5.2

2

Hy3 reportedly wins on most benchmark categories against GLM-5.2 despite running at roughly half the parameter size, a meaningful efficiency claim if it holds up independently

3

Coding remains the one category where GLM-5.2 still outperforms Hy3, suggesting Chinese open-weight labs are converging on different specialization trade-offs rather than one model dominating every category

4

The release adds to a fast-moving 2026 open-weight competitive field that also includes Mistral's upcoming July model, intensifying pressure on closed-weight labs' pricing power

TC
The VC Read · Trace's TakeTrace Cohen

A model winning most benchmarks at half the size of its nearest open-weight competitor is the kind of efficiency claim that, if it holds up independently, matters more to enterprise buyers than another point of raw capability -- inference cost compounds every single day a model is in production. Founders building on open-weight infrastructure should treat this as one more data point that the China-vs-China open-weight race is now as consequential to watch as the US closed-weight race.

Tencent released its Hy3 model under an Apache 2.0 license this week, directly targeting Zhipu AI's GLM-5.2 as its benchmark comparison point and winning most evaluated categories despite running at roughly half GLM-5.2's parameter size, according to VentureBeat's July 6 coverage.

The headline claim is efficiency: if Hy3 genuinely matches or beats a larger competitor's performance across most tasks at half the parameter count, that's a meaningful inference-cost advantage for any enterprise or developer choosing between the two -- smaller models are cheaper to run and faster to serve at scale, all else equal. The one clear exception is coding, where GLM-5.2 reportedly still holds an edge, suggesting the two Chinese labs are making different architectural trade-offs rather than one clearly leapfrogging the other across the board.

The release adds another entrant to a Chinese open-weight AI field that's moved fast in 2026, alongside DeepSeek's continued relevance and DeepSeek's broader influence on how open-weight distribution strategy works internationally. It also lands the same week Mistral is preparing its own open-weight model with July early access, meaning the open-weight competitive field spans both Chinese and European labs simultaneously, each using openness as a distribution wedge against closed-weight incumbents OpenAI and Anthropic.

For enterprises evaluating model choice, an Apache-licensed model winning most benchmarks at half the size of a comparable open competitor is a genuinely useful signal, though independent benchmark verification (rather than the releasing lab's own claims) will matter more than the initial announcement, given how frequently benchmark methodology varies between labs' self-reported comparisons.

What to watch: whether independent evaluators reproduce Tencent's benchmark claims, and whether Hy3's efficiency advantage changes enterprise adoption patterns in coding-adjacent versus general-purpose use cases given its one disclosed weakness.

ShareXLinkedInEmail

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$3,999 local AI workstation

AMD's $4K Ryzen AI Halo Bets on Owning Local Inference

AMD launched the $3,999 Ryzen AI Halo, a compact workstation for running AI models up to 200B parameters locally, betting enterprises want to own inference hardware over rising cloud API costs.

AIRegulatory oversight gap admission

UK Regulator Warns of AI 'Arms Race' in Financial Services

A UK financial regulator warned it is in an 'arms race' just to keep pace with how quickly banks and financial firms are deploying AI, highlighting a widening gap between AI adoption speed and regulatory oversight capacity.

AIDefault opt-in AI training data use

If You Use Google, You're Training Its AI. Here's the Opt-Out.

TechCrunch published a guide reminding users Google uses activity data by default to train its AI, and detailed the settings needed to opt out.

@Trace_Cohen·t@nyvp.com