Nathan Lambert, AI2 post-training lead, says FireworksAI's GLM model integrates with Claude Code in five minutes

@natolambert @FireworksAI_HQ right??

@natolambert @FireworksAI_HQ Thank you! I love using GLM 5.2 with Claude Code.

We've made it one-click experience with FireConnect.

We're also rolling out a change Monday to improve caching for Claude Code requests. More coming soon!!

https://github.com/fw-ai/fireconnect

@xlr8harder @FireworksAI_HQ I think what we're feeling as "better" is faster than opus.

@natolambert @FireworksAI_HQ i admit my memory of opus 4.5 is hazy at this point, but it is succeeding fairly well on some tasks I think would have been hard for opus 4.5, though it's hard to say for sure without testing, which I suppose I could do.

@natolambert @FireworksAI_HQ Feels that it has crossed the same threshold as opus 4.5 back in december, and is maybe even a little better than that, in the tasks i've tried so far.

@natolambert @FireworksAI_HQ not sure if it matters but i'm also trying it in noumena's code harness.

@natolambert @FireworksAI_HQ besides the capabilities and being open source i also agree with this, its very easy to setup and ig that gives them a lot of consumer value

@natolambert @FireworksAI_HQ are you using it for subagents or as a main agent in claude code?

@natolambert @FireworksAI_HQ Had been debating giving it a try because I wasn't sure which provider and harness but this helped me seal the deal

@natolambert @FireworksAI_HQ what is tps?

@natolambert @FireworksAI_HQ my last provider switch: half a day just on auth config

@natolambert @FireworksAI_HQ the claude code integration in 5 min is actually the bigger flex here

is tooling the real moat for these models now?

@natolambert @FireworksAI_HQ It feels opus level to me, at least opus from last December. But I’m hearing it burns a lot of tokens per task.

The next big task for open source models is to get task completion super token efficient.

@natolambert @FireworksAI_HQ Five-minute setup matters more than another benchmark point for adoption. If an open model can drop into Claude Code with almost no plumbing, people will compare it on real work instead of leaderboard prompts. The gateway layer is becoming part of the model product.

@natolambert @FireworksAI_HQ Did you try Fireconnect?

@natolambert @FireworksAI_HQ The practical frontier right now is less about which lab wins and more about how quickly you can swap models into an established agent harness without breaking the surrounding eval and memory layer.

@natolambert @FireworksAI_HQ This feels small, but small behavior changes are usually where bigger shifts start.

@natolambert @FireworksAI_HQ 五分钟跑起来这点很爽了

@natolambert @FireworksAI_HQ First hour is the easy hour — setup + greenfield flatter any decent model. The real tell is hour six: does GLM hold up once the context is full and the task spans files, or do small reasoning slips start compounding? That's where my loops (and my bills) usually find the gap.

@natolambert @FireworksAI_HQ Setup friction is underrated. 5min vs 30min onboarding is the difference between trying and not.