What is GLM-5.2? Another open-source Chinese AI model has Silicon Valley's attention.

tech

What is GLM-5.2? Another open-source Chinese AI model has Silicon Valley's attention.

GLM-5.2, a new open-source Chinese AI model for long-form coding tasks, is turning heads among Silicon Valley's AI elite.

TL;DR

GLM-5.2 is a new open-source AI model developed in China by z.AI.
It is designed for long coding tasks and agentic workflows and boasts a 1 million token context window.
Industry leaders have praised its coding capabilities, with some calling it a viable 'daily driver'.
As an open-source model, it can be freely downloaded and modified, unlike most US frontier models which are closed.
This open-source nature could allow GLM-5.2 to capture a larger market share, potentially challenging US AI supremacy.
China's advancement in AI is seen as a response to US efforts to maintain its edge through chip restrictions.

Continue reading the original article

clem 🤗 reposted

Graham Neubig

OK, I tried GLM-5.2 and this is a good model.

Probably the first model good enough to eschew closed models from your workflow entirely (except if you need vision).

I know this won't run on your laptop, but what are the best current vllm/sglang serving recipes?

clem 🤗 reposted

Harry Stebbings

“We have a crisis of open source models in the Western world.

Outside of China, there are no good open source models. We don't even have any in the US now.

The talent, the capital and the focus to be best-in-class at pre-training, mid-training, post-training is an extremely scarce skillset.

The answer for a lot of countries may just be to take an open weights model from China, post-train or fine-tune their own version, and have that be what they start from." @EverettRandle

Why does the West have such poor open source models and is it a national security threat to rely on open-source Chinese models @demishassabis @ylecun @aidangomez @lilianweng @janleike

clem 🤗 reposted

Harrison Kinsley

Replying to post

GLM 5.2 quantized GGUFs: https://t.co/IVIW3VpF5T GLM 5.2 on Open Router: https://t.co/pSuedaqJtM Minion coding agent: https://t.co/vpzWo2UfXo A decent consolidation of models/benchmarks: https://t.co/WOhkOka8zs DeepSWE: https://t.co/1wPtE4P0di Throwback to ChatGLM paper a few

clem 🤗 reposted

Patrick C Toulme

I ran GLM 5.2 with OpenCode harness against Claude Opus this week deployed locally.

Bottom line: It is a real frontier coding model and insanely good for the price (free). Open source model + open source harness + local serving on my own chips is an amazing value proposition.

Some Notes:

1. Tool calling is very good — spun up nested subagents on its own, multiple levels deep

2. Very good at research and planning including long range plans

3. It built a cell based terminal renderer at near Opus quality. I still lean Claude, but most people couldn't tell the outputs apart

4. Opus wins on oneshotting and reading my intent without me explicitly telling it

5. GLM 5.2 is more than enough intelligence for most F500 work IMO

6. GLM 5.2 is good enough to hill climb RL with and to drive further AI development / next generation GLM model. Progress will be much faster for their RL from here.

7. Running my own endpoint = permanent fast mode

8. It wastes thinking tokens writing code in the reasoning block

At this point I would consider GLM 5.2 a true frontier coding model. Getting to this point in coding quality was the hardest part IMO. They will progress quickly from here in RL.

clem 🤗 reposted

Zixuan Li

GLM-5.2 has been "stuck" at No.2 on Hugging Face Trending for three days, but I'm thrilled to have connected with the creator behind the No.1 project this afternoon. It's been amazing to see open-source work resonating with so many people. https://t.co/fRb4UXvtvV

Aravind Srinivas reposted

Bread🍞

GLM is pretty solid — it gets about an 80% pass rate on our internal financial benchmark. By contrast, DeepSeek v4, Kimi and MiniMax are below 5%.

(Considered Opus 4.8 as baseline & judger)

clem 🤗 reposted

Itamar Golan 🤓

Hot take:

GLM 5.2 might be the first open/public model that actually changes the enterprise AI cost equation.

I played with it for a few hours today after friends told me to stop ignoring it.

I expected the usual open model experience:

“Cool, but clearly not frontier.”

That’s not what I felt.

Across a bunch of different tasks, it was the first public open model that didn’t feel obviously worse than the best closed models. Not perfect. Not fully benchmarked. But shockingly capable.

The first open model that made me think:

“Wait, this could actually replace a meaningful chunk of frontier model usage.”

That matters because the economics are wild.

Yes, serving it properly is expensive. Something like 8 Nvidia H200s, around $400K to buy or $20K/month to rent.

But many enterprises are spending millions per month on Anthropic/OpenAI.

If GLM 5.2 is even close enough for 30-50% of workloads, that’s not a model launch.

That’s a market disruption.

clem 🤗 reposted

⚡AI Search⚡

GLM 5.2 continues to impress me.

Here's its result on Vending Bench, which measures an AI's performance on running a business over a long time.

GLM 5.2 came in second, while costing less than HALF of Opus. https://t.co/IXOgXyQMJM

Aravind Srinivas reposted

the tiny corp

I have on good authority that GLM 5.2 is running at 120 tok/s across two networked Blackwell tinyboxes. $150k and that setup can be yours, either 2x tinybox or 1x tinybox pro. Never pay the cloud again.

Aravind Srinivas

GLM is the kind of model that revives serious interest in open source AI. It passes the blind test relative to the frontier models on the median production grade knowledge worker task. It’s affordable to serve. And is a sub trillion parameter model, meaning it has a lot of potential to go beyond matching the frontier at the median level of difficulty to also doing it for the long tail. Plenty to look forward to!

clem 🤗 reposted

Onur Solmaz

I created an LLM leaderboard based on Hugging Face download and like counts, grouped, filtered and time-averaged. Top 5 downloads is shared by @Alibaba_Qwen and @googlegemma 👑🤝👑

Top 5 likes, on the other hand also includes @deepseek_ai V4 Pro 👑

Even @OpenAI makes it to #8 top downloads with gpt-oss-20b 👑

qwen3-6-35b-a3b is the second most CIRCULATED LLM of this year, with an average of 21 million downloads per month, since the day it was released 2 months ago 📈📈📈

Despite first place belonging to 8mo old qwen3-vl-2b-instruct, the highlight belongs to the mid-sized MoE model, which has hit a size/performance sweet spot so hard that it absolutely 💥 SHATTERED 💥 Hugging Face leaderboards in the 2 months since it has launched

qwen3-6-35b-a3b is followed closely by its dense sibling 27b --- and then the mid-sized gemma 4 models 26b-a4b and 31b

Note that a model's distribution is inversely proportional to its size, but not strictly! Usefulness plays a factor as well, since gemma 4 26b-a4b is being downloaded more than the smaller gemma 4 e4b

I created this leaderboard because Hugging Face's all time highest downloads and likes did not give me enough information about what is really popular, neither today, nor all-time. I wanted something in between

How do I calculate this ranking?

- Get models that with n_downloads >= 100k

- Exclude models older than 1 year

- Deduplicate and group quantizations and variants of the same model based on slug prefix heuristics

- For each group, sum up total downloads of all time

- Sort by descending total_downloads / age = average_downloads_per_day (can also sort w.r. to likes per month)

- Repeat every day to get the most up to date ranking

More info and source on the leaderboard page, hosted on a Hugging Face space: https://t.co/m6SapLztCC

This is a work in progress, please reply below if you see a model that should be there is missing, or any other mistakes

clem 🤗 reposted

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)

Good guy DeepSeek gives us accelerated models

The most interesting one here is Gemma4-12B, I presume vision included. Might be the best local model in its weight class now, by some margin

Qwen 3.5 not included because DS[park] doesn't do linear attention I guess https://t.co/lKtcGBQH3w