GLM 5.2 and the gap between an LLM and a VLM

A top open-weight model that rivals the closed frontier on coding - and cannot see a thing. What it takes to give a model eyes, and why even an open-weights leader keeps vision behind a closed API.

June 20, 2026 9 min read
LLM into a VLM Vision kept separate Opus and Codex on images Which one to use References

Zhipu shipped GLM 5.2 on June 13, 2026, weights under an MIT license, no regional locks, free to download and self-host. Within days it was the model everyone in the open-weight world was arguing about. Artificial Analysis put it at 51 on its Intelligence Index v4.1, level with Claude Opus 4.8 and clear of the other open models like DeepSeek V4 Pro and Kimi K2.6. On coding it looks even stronger. It scored 62.1 on SWE-bench Pro, ahead of the roughly 58.6 reported for GPT-5.5 (the model now behind Codex), and 81.0 on Terminal-Bench 2.1, only a few points under Opus 4.8's 85.0. You can see why people started saying it beats the closed frontier.

A high-end camera lens in dramatic warm light against deep navy shadow, light refracting through the glass - an eye about to open
0
Intelligence Index v4.1
Level with Opus 4.8
0
SWE-bench Pro
Ahead of GPT-5.5's 58.6
0
Terminal-Bench 2.1
Just under Opus 4.8's 85.0

The fuller picture is less tidy. On hard reasoning GLM 5.2 still trails, sitting behind both Opus 4.8 and Gemini 3.1 Pro on Humanity's Last Exam and behind the top closed models on GPQA-Diamond. So "better than Opus 4.8 and Codex" holds on a couple of coding axes and runs optimistic on the rest.

Where it leads

A couple of coding axes

SWE-bench Pro and Terminal-Bench put it level with or ahead of models behind the closed coding frontier, and the Intelligence Index lands it with Opus 4.8.

Where it trails

Hard reasoning

It sits behind Opus 4.8 and Gemini 3.1 Pro on Humanity's Last Exam, and behind the top closed models on GPQA-Diamond.

But there is a more basic limit the leaderboard framing skips right past: GLM 5.2 is text only. Paste a screenshot, a photo of a stack trace, a Figma export, a chart, and the model has nothing to work with. For something being pitched as a coding agent, that is a real hole, because a large share of real engineering input shows up as pixels.

The hole in a text-only coding agent

A screenshot of a broken UI. A photo of a stack trace on a second monitor. A Figma export. A chart you want read back. To GLM 5.2 these are all the same thing: nothing. A large share of real engineering input shows up as pixels, and a text-only model cannot see any of it.

GLM 5.2 is a good LLM. The harder questions are what it would take to give a model like it eyes, and why the labs that build these models keep vision separate from text.

What actually changes when an LLM becomes a VLM

A vision language model is less a different species of model than an LLM with extra parts bolted onto the front. The mainstream recipe has three pieces.

Vision encoder

Almost always a Vision Transformer from the CLIP or SigLIP family, pretrained on huge image-caption sets. It chops an image into patches and returns a grid of embeddings.

Projector

A connector that maps the encoder's vector space to the LLM's. A small two-layer MLP (LLaVA), or cross-attention reaching into image features (BLIP-2 Q-Former, Llama 3.2 Vision). Out come "visual tokens".

LLM backbone

The model you already have. It keeps doing next-token prediction. The only change: the sequence now starts with tokens that came from an image instead of a tokenizer.

Input

Image

A screenshot, photo, chart or mockup.

raw pixels
1 · Vision encoder

ViT (CLIP / SigLIP)

Patches → a grid of embeddings.

out: ~1408-d vectors Frozen
2 · Projector

MLP or Q-Former

Maps the encoder space into the LLM's.

1408-d → 4096-d Trained
3 · LLM backbone

Next-token prediction

Predicts over the combined sequence.

in: 4096-d tokens Mostly frozen
one sequence the backbone predicts over
imgimgimgimg theerrorisonline

You do not train this from scratch. The standard path is to take a pretrained encoder and a pretrained LLM, freeze most of both, and train the projector to align the two on image-caption and interleaved image-text data. After that alignment stage comes supervised instruction tuning, which is what teaches the model to answer questions about an image rather than just caption it. Plenty of teams unfreeze the top few LLM layers during this phase and leave the vision encoder frozen the entire time, which cuts the training cost dramatically while keeping accuracy.

STAGE 1

Align the projector

Freeze the encoder and most of the LLM; train the projector on image-caption and interleaved data.

STAGE 2

Instruction tune

Teach it to answer questions about an image, not just caption it. Often unfreeze the top few LLM layers here.

RESULT

A working VLM

Encoder stays frozen throughout, which cuts training cost sharply while keeping accuracy.

The line item people underestimate is token cost. One image can blow up into far more tokens than the prompt wrapped around it, sometimes by more than an order of magnitude. That is the whole reason high-resolution handling, tiling, and visual-token reduction are their own active research areas, and it is why "just add vision" is never actually just adding vision. You are paying for it on every forward pass.

Text prompt ~ a couple hundred
One image 10x or more

paid on every forward pass

Zhipu keeps vision in a separate model

GLM 5.2 being text only is not an oversight. It is a product split. Zhipu runs a separate vision line, the GLM-V series, and the split maps almost exactly onto the architecture above.

Open · self-hostable

The text model

GLM 5.2: MIT weights, no regional locks, free to download and run on your own hardware. The strong, open capability.

Closed · API only

The vision model

GLM-5V-Turbo: the same text core plus a CogViT encoder and a round of multimodal training. Closed-source, rented through an API, not runnable on your own hardware.

The open vision model is GLM-4.5V. It is built on the GLM-4.5-Air text base, 106 billion total parameters with 12 billion active, with a vision transformer attached, 3D-convolution handling for video, and interpolation tricks so it can cope with high-resolution and odd aspect-ratio images. In the current generation the vision model is GLM-5V-Turbo, which is closed-source and API-only. It takes the same text execution core as the GLM-5-Turbo text model, then adds a CogViT vision encoder and a round of multimodal training on top. That is the LLM-to-VLM path described above, walked in production by the lab itself.

The pattern

The strong, open, self-hostable model is the text one. Vision lives in a separate model you rent through an API and cannot run yourself. Even a lab that leads on open weights treats vision as a harder, more guarded capability than text.

Where Opus and Codex actually stand on images

Vision in these coding-focused frontier models still trails their text reasoning. That gap is real, but it is narrowing quickly, and for the cases that come up most often the models are already useful.

GPT-5.2-Codex

Stronger vision in the loop

Called out by OpenAI for reading screenshots, technical diagrams, charts and UI surfaces during a coding session, and turning design mockups into working prototypes.

GPT-5.3-Codex

Vision drives computer use

Leaned on vision to drive computer-use tasks - the OSWorld style of operating a real desktop by looking at what is on the screen.

GPT-5.4

Native computer use

Shipped native computer use as a headline feature. Claude's Opus line takes images too, and follows the same curve: rough a year ago, dependable now.

Do not hand one of these models a dense, label-heavy engineering schematic and expect clean extraction. But a screenshot of a broken UI, a chart you want read back, or a mock you want built will work - and it works better with each release. When you choose a model, weigh where vision is heading, not just where it sits today.

Which one to use

GLM 5.2 is a strong text model, and the thing it skips is the one that would let a coding model look at your screen the way a teammate could. Adding vision is a known process: a vision encoder, a projector, and alignment training. But it carries real compute cost and real engineering difficulty, which is why Zhipu ships its own vision model as a paid, closed API instead of folding it into the open download.

For picking a model, it comes down to a trade. GLM 5.2 gives you a top open model you can run on your own hardware. The closed models you might otherwise use can already read a screenshot, a chart, or a mockup you paste in. If your work involves images, that usually settles it.

“

GLM 5.2 gives you a top open model you can run yourself. The closed ones can already see. If your work involves images, that usually settles it.

References

Choosing the model is the easy part

Open or closed, text or vision - the model is one decision among many. We help teams pick the right one for the job and stand it up so it keeps working in production. Get a free expert review of your AI idea or roadmap, read by our Chief AI Officer.

Get your free expert report