I am, performance is actually ok if I use anything under 12B data points on reasoning models. In terms of code models, I like Qwencoder for speed on GPT4ALL but I have been using Codeistral (Mistral) 22B with good results on LM Studio. You need to install CUDA for Nvidia cards to get best results.

I tested Qwen Coder Next — a free, local coding model that runs entirely on your own hardware. No tokens, no monthly fees. Here's what it can (and can't) do 👇
youtu.be/jDeeoHSc2kw
#AI #LocalLLM #QwenCoder #CodingTools

I'm a huge huge proponent in open source models we can run on our own hardware. The QwenCoder series is doing great, and can run on my laptop (which, admittedly, is an absolute unit), and related models on inexpensive cloud GPUs.

I foresee (and would like to help) an explosion of varieties of these

Well the initial vibe check isn’t great but I need a larger sample set. QwenCoder 2.5 did an okay zero shot but since it's not a thinking model I followed up with a prompt of "are there any bugs in the harness?" and it went off the rails producing incoherent output

New #devlog is up for Montana #Pygame Workbench 🎮🛠️

Added #AI Coach workflow wrappers for #Gemma, #QwenCoder, and #DeepSeekCoder. Cleaned Up Cloud Hookups.

Better Explain → Plan → Patch → Audit → Repair behavior, with reviewable Accept / Deny edits.

Devlog:
montana-jay.itch.io/montana-pyga...

QwenCoder is kicking Gemini-Cli bad. Does the job real well and sometimes better than Gemini. Chinese models are crushing it. Only thing pending is hardware and China might be catching up soon. Hardware is where India should be putting its effort in and not just the models, but alas

"The hottest new programming language is English," @karpathy.bsky.social.
Open-source models are becoming extremely good at generating code.
Example: QwenCoder-32B performs on par with GPT4o.
Yet, devs are still not taking enough advantage of powerful OS coding LLMs.
Let's change that!
code.lycee.ai

In case you weren't aware, you can download smaller versions of Gemma, Llama, QwenCoder, etc. You can set up chats in LM Studio with them, or you can use them in code editors with extensions like Continue. You can make API calls to them with ollama.

2/n

and even several larger ones. We show the generality of our approach on two model families (DeepSeekCoder and QwenCoder), compare against other fine-tuning approaches, and demonstrate robustness by showing retention of code generation abilities [6/7 of https://arxiv.org/abs/2503.03656v1]

projection onto the base model to avoid overfitting. Using our approach, we obtain a new series of models NextCoder (adapted from QwenCoder-2.5) that achieves strong results on five code-editing benchmarks, outperforming comparable size models [5/7 of https://arxiv.org/abs/2503.03656v1]

This context retrieval trick addresses context related problems and could boost models productivity. It's being tested on a relatively small 1.8B QwenCoder 2.5, and looks very promising to me

huggingface.co/blog/codelio...

@swyx https://x.com/swyx/status/1976743791437693270 #x-swyx

Quote: https://x.com/shydev69/status/1976641232622453045

Do I know anyone affiliated with GLM? we are looking for one last speaker for AIE CODE and a GLM/QwenCoder/Kimi researcher is probably the rig...