←back to thread

GLM 4.5 with Claude Code

(docs.z.ai)
210 points vincirufus | 3 comments | 06 Sep 25 00:45 UTC | HN request time: 0.615s | source
Show context
chisleu ◴[06 Sep 25 02:07 UTC] No.45145959[source]▶
I've been using GLM 4.5 and GLM 4.5 Air for a while now. The Air model is light enough to run on a macbook pro and is useful for Cline. I can run the full GLM model on my Mac Studio, but the TPS is so slow that it's only useful for chatting. So I hooked up with openrouter to try but didn't have the same success. Any of the open weight models I try with open router give sub standard results. I get better results from Qwen 3 coder 30b a3b locally than I get from Qwen 3 Coder 480b through open router.

I'm really concerned that some of the providers are using quantized versions of the models so they can run more models per card and larger batches of inference.

replies(3): >>45145970 #>>45146999 #>>45149106 #
1. KronisLV ◴[06 Sep 25 06:05 UTC] No.45146999[source]▶
> I get better results from Qwen 3 coder 30b a3b locally than I get from Qwen 3 Coder 480b through open router. I'm really concerned that some of the providers are using quantized versions of the models so they can run more models per card and larger batches of inference.

This doesn't match my experience precisely, but I've definitely had cases where some of the providers had consistently worse output for the same model than others, the solution there was to figure out which ones those are and to denylist them in the UI.

As for quantized versions, you can check it for each model and provider, for example: https://openrouter.ai/qwen/qwen3-coder/providers

You can see that these providers run FP4 versions:

  * DeepInfra (Turbo)
And these providers run FP8 versions:

  * Chutes
  * GMICloud
  * NovitaAI
  * Baseten
  * Parasail
  * Nebius AI Studio
  * AtlasCloud
  * Targon
  * Together
  * Hyperbolic
  * Cerebras
I will say that it's not all bad and my experience with FP8 output has been pretty decent, especially when I need something done quickly and choose to use Cerebras - provided their service isn't overloaded, their TPS is really, really good.

You can also request specific precision on a per request basis: https://openrouter.ai/docs/features/provider-routing#quantiz... (or just make a custom preset)

replies(1): >>45147398 #
2. snthpy ◴[06 Sep 25 07:46 UTC] No.45147398[source]▶
Interesting. Thanks for sharing. What about qwen3-coder on Cerebras? I'm happy to pay the $50 for the speed as long as results are good. How does it compare with glm-4.5?
replies(1): >>45147831 #
3. KronisLV ◴[06 Sep 25 09:25 UTC] No.45147831[source]▶
I wish that Cerebras had a direct pay per use API option instead of pushing you towards OpenRouter and HuggingFace (the former sometimes throws 429, so either the speed is great, or there is no speed): https://www.cerebras.ai/pricing but I imagine that for most folks their subscription would be more than enough!

As for how Qwen3 Coder performs, there's always SWE-bench: https://www.swebench.com/

By the numbers:

  * it sits between Gemini 2.5 Pro and GPT-5 mini
  * it beats out Kimi K2 and the older Claude Sonnet 3.7
  * but loses out to Claude Sonnet 4 and GPT-5
Personally, I find it sufficient for most tasks (from recommendations and questions to as close to vibe coding as I get) on a technical level. GLM 4.5 isn't on the site at the time of writing this, but they should match one another pretty closely. Feeling wise, I still very much prefer Sonnet 4 to everything else, but it's both expensive and way slower than Cerebras (not even close).

Update: also seems like the Growth plan on their page says "Starting from 1500 USD / month" which is a bit silly when the new cheapest subscription is 50 USD / month.