VC
Value Add VC
⚡HomePulse⚡Helpful Apps📝Blog
← Value Add PulseAIHalf-size open model, Apache 2.0

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

Tencent released Hy3, an Apache-licensed open-weight model roughly half the parameter count of Zhipu's GLM-5.2, claiming wins across most benchmark categories except coding tasks -- intensifying the open-model race among Chinese AI labs.

Apache 2.0
License
~Half the parameters
Size vs. GLM-5.2
Coding benchmarks
Exception
TC
Trace Cohen
Early-stage VC & angel · Founder, New York Venture Partners
July 6, 2026
2 min read
ShareXLinkedInEmail
THE RUNDOWN
1

Hy3's Apache 2.0 license makes it freely usable for commercial deployment without the licensing restrictions some competing open-weight models carry, a meaningful factor for enterprise adoption

2

Beating a larger competitor model at half the parameter count would represent a genuine efficiency gain, translating directly into lower inference cost for anyone deploying it

3

The one exception -- coding tasks, where GLM-5.2 reportedly still wins -- suggests specialized post-training or data curation still meaningfully matters even as general efficiency improves

4

This continues a 2026 pattern of Chinese AI labs (Tencent, Zhipu/GLM, Alibaba's Qwen, DeepSeek) competing intensely on open-weight model efficiency, a category where Western labs like OpenAI and Anthropic have been comparatively less active

TC
The VC Read · Trace's TakeTrace Cohen

Every time a smaller open model claims to beat a larger one, the actual founder-relevant question is inference cost per task, not the benchmark leaderboard position -- and if Hy3's efficiency claim holds, that's real margin back for anyone building on open weights. The coding exception is the tell that specialized post-training still has a moat, at least for now; that's exactly the kind of narrow wedge a vertical AI coding startup should be watching closest, not the general-purpose leaderboard.

Tencent released Hy3, a new open-weight AI model licensed under the permissive Apache 2.0 license, claiming it beats Zhipu AI's GLM-5.2 across most benchmark categories despite having roughly half the parameter count -- with coding tasks as the notable exception, where GLM-5.2 reportedly retains an edge.

The efficiency claim, if it holds up under independent testing, matters because parameter count correlates directly with inference cost: a model that matches or beats a larger rival's general capability at half the size translates into meaningfully lower compute cost for anyone deploying it at scale, a increasingly important differentiator as enterprises grow more cost-conscious about AI deployment amid rising concern over usage-based API pricing.

The Apache 2.0 license is a deliberate choice that removes many of the commercial-use restrictions some competing open-weight models carry, positioning Hy3 for broader enterprise and developer adoption without the legal ambiguity that has occasionally complicated deployment of other open models with more restrictive custom licenses.

“That divergence has made the open-weight model race an increasingly Chinese-lab-dominated category globally.”

The coding-task exception is a meaningful data point on its own: it suggests that even as general-purpose efficiency improves rapidly, specialized capabilities like code generation still benefit from targeted data curation and post-training investment that a smaller, more general-purpose model hasn't fully replicated -- a pattern consistent with how coding-specific models and fine-tunes have carved out a persistent niche even as general frontier models improve.

This release continues an intensifying competitive dynamic among Chinese AI labs specifically: Tencent's Hy3, Zhipu's GLM series, Alibaba's Qwen family, and DeepSeek have all been iterating rapidly on open-weight models throughout 2025 and 2026, a category where Western frontier labs like OpenAI and Anthropic have generally been less active, preferring closed, API-only deployment for their most capable models. That divergence has made the open-weight model race an increasingly Chinese-lab-dominated category globally.

For enterprises and developers evaluating model choice, Hy3's combination of permissive licensing and claimed efficiency gains adds another credible option to an already crowded open-weight landscape, intensifying the "good enough and much cheaper" pressure that open models are increasingly putting on closed frontier-lab pricing.

The bear case: benchmark claims from any single lab's own release announcement warrant independent verification before being taken at face value, and "wins everywhere except coding" is still a meaningful gap given how central coding capability has become to enterprise AI adoption and developer-tooling use cases specifically.

What to watch: independent benchmark verification of Hy3's claimed performance advantages, whether Tencent follows with a coding-specialized variant to close the remaining gap with GLM-5.2, and whether this accelerates pricing pressure on closed frontier-lab APIs as open alternatives keep improving efficiency.

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