What is the best coding LLM per dollar?Not the "best" (probably Claude) but the best value?
>>109154216GLM5.2
Local LLM, it's literally free, can't beat that bang for buck./thread/sage
probably still the subs from openai and anthrodeepseek v4 pro probably if it's just api costglm is an awkward middle ground because it's decent but people complain that the service from their sub is terriblethe harder your task the better off you are with a sub to the western frontier - it's just a lot less hassle and wasted timefor example>>109154296in theory local is the best because token cost is very low, but unless you can run ds or glm locally, the braindead models are going to take hours of tardwrangling to do things current gpt and claude on low reasoning efforts will just one shot in a couple of minutes
>>109154296Sure, if you have the right hardware. I don't know about you but I can't afford to have 512 GB of RAM to run GLM-5.2.
>>109154296hardware costs money
>>109154367qwen3.6 isn't bad
>>109154216Qwen3.6 27B running on Arc B70 32GB
>>109154216Deepseek v4 flash as a subagent and glm 5.2 as the senior dev assign tasks and check its work.
>>109154216Depends how you measure "best".Most slop per hour?
>>109154296Nigger your 20k context window can't even read a single spec sheet describing the database architecture.
>>109154455Why would it need to? Surely you are not a mouthbreathing idiot incapable of instructing the LLM what to do step by step?
>>109154328Antirez (the guy who created redis) has been doing some incredible work optimizing deepseek and glm 5.2 for local. We're so damn close to local being "good enough" the hardware is just constrained by this retarded memory shortage.
>>109154469yeah that work is great, but the memory shortage is going to get worse for the foreseeable future and nvidia and amd aren't going to put out 128gb gpu's at reasonable prices even if memory prices drop to rock botton
>>109154463>needs llms to build babies first bash script>thinks his opinions on software architecture are relevant
>>109154486Literally the opposite of what i said. Maybe utilize your llm context to understand a post next time.
>>109154216Deepsneed 4 pro is pretty much free which is hard to beat in terms of value
>>109154216the one you have with you
>>109154492>doesn't even understand the point>thinks his opinions are relevant
>>109154290>>109154328I'm tempted to try the z.ai Pro tier since its a third cheaper than Claude Pro but its still not as much as I would have hoped(Unless the context window / usage limit is way bigger)>>109154429What platform / tool do you use to do that?Or are you using v4 flash locally and glm with a host and using both manually?
>>109154216I would guess OpenAI ChatGPT Plus. You get a lot of usage for the low tier sub. Even on medium reasoning effort it outperforms the Chinese models by a lot and is on the same level on coding as Opus.When you have to pay API GLM5.2 is not even cheaper than GPT 5.5 on medium, OpenAI models work much faster and are far more token efficient.You can get a lot of inference though on GLM 5.2 on the Opencode Go sub. $10 for $60 of inference on open-weight models, where you also can delegate tasks to Deepseek models or M3, which are dirt cheap. Vision also works decently with MiMo and is cheap.Another option to try is the Grok subscription. It provides Composer 2.5 from Cursor and is also a very capable coding model, and you get a lot of inference. They also want to publish a new more capable model in July, but have to see how good that is going to be.
