GLM 5.1 Benchmarks
Posted by Fantastic-Emu-3819@reddit | LocalLLaMA | View on Reddit | 25 comments
GLM 5.1
Posted by Fantastic-Emu-3819@reddit | LocalLLaMA | View on Reddit | 25 comments
GLM 5.1
pmttyji@reddit
I think GLM-5.1 set the bar high for DeepSeekV4.
Yes-Scale-9723@reddit
i'm also waiting for the new version of deepseek. currently it has an outstanding value for money
power97992@reddit
As long v4 is just as good as 5.1 but 8x cheaper , it will be great!
SourceCodeplz@reddit
I've used GLM-5 and it is fantastic for my usage OOP PHP. Same as Sonnet 4.5 really for me.
Radiant_Hair_2739@reddit
wow, now I have the local GPT-5.4 on my local server PC with Epyc with 512gb RAM DDR4, GLM-5 given pp = 110 t/s with tg = 5.5 t/s, thanks!
Yes-Scale-9723@reddit
That's great but for coding agents 5.5 t/s is really slow, it will take hours to complete a typical 50k token task
Radiant_Hair_2739@reddit
it doesn''t matter. Because if I try to run Opus 4.6 for agentic task using API, for example here I should pay almost 20$ per difficult task which I can complete almot free using local GLM:
Yes-Scale-9723@reddit
well in that case it's a great solution, i forgot how expensive are those models
Caffdy@reddit
8 or 12 channels?
pmttyji@reddit
Nice. Hope you're using optimized llama.cpp command. Also ik_llama
Specter_Origin@reddit
I hope it has faster inference speed than last one…
NandaVegg@reddit
It's somehow much faster than 5 in all inference providers in spite of the same-ish architecture (fp8).
Remote_Rutabaga3963@reddit
Meh. Faster than at launch sure
atape_1@reddit
Coding benchmarks are absolutely wild.
-dysangel-@reddit
I've been using it for coding the last few weeks. It's good!
pigeon57434@reddit
the most important thing for me is is this model more CoT efficient because glm models always seem to think for like 97 years for me and im using it on zhipus official website so its not even a local hosting skill issue
Edzomatic@reddit
From my very limited testing it does indeed think less, and the final output also has less AI fluff
Xisrr1@reddit
It's been like that since GLM 5 already, now even more efficient
EndlessZone123@reddit
no vision still :(
kaggleqrdl@reddit
AHAHA GLM 5.1 announces SOTA and Anthropic comes back with .. a model you can't use. LOL. PANIC
LittleYouth4954@reddit
I've been using glm 5.1, 5-turbo and 5v for a week now and they are amazing. I am also impressed by qwen 3.6.
LegacyRemaster@reddit
Unfortunately, to make it run at at least 20 tokens/sec on 192 GB vram I would have to limit myself to IQ1... So a few percentage points above minimax or qwen are almost certainly lost in quantization.
Makers7886@reddit
Agreed, imo the best model right now for 192gb vram is qwen 3.5 122b FP8 via vLLM. Over 80 t/s solid and 220-240 with 6+ concurrent and 200k context. Every time I "stretch" for a large model I lose the speed, concurrency, and context in exchange for "checking out the big dog" which is simply not usable for real purposes or at least feels unusable because of all the cons.
ambient_temp_xeno@reddit
It's 1.9% better than gemma 4 31b on GPQA-Diamond.
I guess I'll use all that ram for gemma SWA checkpoints because I'm guessing I'd lose that 1.9% advantage running GLM 5.1 in IQ1.
Ok-Measurement-1575@reddit
So... Minimax is basically the best pound for pound LLM right now?
Where dem weights at? :D