@PatrickToulme: I ran GLM 5.2 with OpenCode harness against Claude Opus this week deployed locally. Bottom line: It is a real frontier …

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.

Agreed. I can see a world in which some customers and enterprises still pay the closed source model premium for the max intelligence, but many enterprises who do not want to spend billions will offer employees self hosted open source models for low cost.

I served it on H100s. I did not measure tok/s but its faster than Claude generations Id say

I’m not a bot. I ran it on H100s

They wil hill climb with RL though. Getting GLM models to this point in agentic coding was the hardest part. It is exponential growth from here assuming they have enough compute.

There is a price quality curve. For example most F500s if they can get Opus tier for almost free will give that to most employees versus Claude Fable.

However there will still be some customers example being quant funds who need the absolute highest intelligence and will pay whatever Anthropic or OpenAI charge

I ran this on 8 H100s. Was fast but honestly still too slow. Will try 8 Blackwells soon.

I am trying to get it to run on TPUs right now as well.

H100. hopefully TPUs soon

Yes 100%

They will gather a large amount of trajectories and RL on positive ones. Just time + compute. There is no secret.

h100x8 did not measure token/s

Because GLM 5.2 has so many positive trajectories.