The best open-weight AI models for coding in 2026, grouped by the hardware you need to run them

The best open-weight AI models for coding today are GLM 5.3, Kimi K3, DeepSeek V4 Pro and the large version of Qwen 3.8. The catch is that almost none of them run on your computer. So I split the ranking by the machine you need to run them: datacenter, multi-GPU workstation, consumer computer. The third tier has a single model: Qwen 3.8 27B, which fits on a 24 GB GPU.

What are the most powerful open-weight models for coding?

The first tier is the most powerful models, with performance getting close to Fable and GPT. They only run on datacenter hardware.

• GLM 5.3 by Z.ai. When I recorded the video, in September 2026, it was the open-weight model with the highest score on Artificial Analysis. To run it well you need multi-GPU servers, for example machines with 8 NVIDIA H200s.
• Kimi K3 by Moonshot AI. Its size is off the charts: 2.8 trillion parameters, 104 billion active per token and about 1.6 TB of weights. To keep it all in memory you’re looking at a machine with 16 NVIDIA H200s.
• Qwen 3.8 (the large version), DeepSeek V4 Pro and Kimi K2.7 Code: all huge models, above one trillion parameters. DeepSeek V4 Pro, for example, has 1.6 trillion, 49 billion of them active.

You can still explore Kimi K3 on a Mac, with some trade-offs: I covered it in how Kimi K3 runs locally on Apple Silicon.

Why is DeepSeek V4 Pro the most interesting on price/performance?

DeepSeek V4 Pro stands out mainly for the ratio between performance and cost. On Artificial Analysis it scores 36 points on the Intelligence Index, against 53 for Fable 5.1. That’s almost 70% of the score of one of the best proprietary models.

Cost changes everything. On the same index, DeepSeek V4 Pro costs $0.67 per task. Fable 5.1 costs $7.63 per task, more than ten times as much.

For many development tasks that 30% quality gap matters. For others it doesn’t, and there the cost difference weighs more than the score.

How do you use these models without a datacenter?

First-tier open-weight models are used through an API. The simplest option I know is OpenRouter: it gives you access to hundreds of models through a single API, and you can switch from one to another without changing your code.

OpenRouter also plugs into coding agents like OpenCode and Pi. That way you can try GLM 5.3 or DeepSeek V4 Pro inside the same workflow you would use with a proprietary model.

Alternatively, you can connect directly to the labs’ own APIs: DeepSeek, Z.ai, Moonshot, Alibaba and the others.

Which open-weight models run on a multi-GPU workstation?

The second tier is smaller models that still need a lot of GPU memory. Here the setups start to become realistic: GPUs rented by the hour, or a multi-GPU workstation.

This tier includes models like DeepSeek V4 Flash and GLM 5.3 Flash. They still have hundreds of billions of parameters, but they only activate a few tens of billions per token. GLM 5.3 Flash, for example, has 320 billion parameters and activates 18 billion.

They are still huge models, but far more efficient than their bigger siblings. The price you pay is lower performance.

Which open-weight model actually runs on a consumer computer?

The only model in this ranking you can truly run on a consumer computer is Qwen 3.8 27B, by Alibaba, with 27 billion parameters.

With 4-bit quantization, which stores the weights at lower precision so they take less memory, Qwen 3.8 27B weighs between 15 and 18 GB. A GPU with 24 GB of VRAM, like an RTX 3090, can run it locally.

Context window and speed of Qwen 3.8 27B running locally depend heavily on the setup: how much memory is left after the weights decides how much context you can keep open.

Does Qwen 3.8 27B really perform like Opus?

No. And this is where you need to be careful when you read "as good as Opus".

On some benchmarks Qwen 3.8 27B really is close to Opus, and in some cases it even beats it. But as soon as you raise the difficulty, the picture changes completely.

Terminal-Bench 3.0 is one of the hardest benchmarks for evaluating an AI agent that has to work autonomously in the terminal. On Terminal-Bench 3.0 Opus 5 scores 42.7%. Qwen 3.8 Max, the bigger sibling of Qwen 3.8 27B, reports 11.3%. Qwen 3.8 27B doesn’t even appear on the public Terminal-Bench 3.0 leaderboard.

It’s impressive that a 27-billion-parameter model, running on a computer like yours or mine, gets results like these. But saying Qwen 3.8 27B performs like Opus is a big stretch.

Which open-weight model should you pick for coding?

It depends on where the model has to run:

1. You want the best performance: GLM 5.3 or Kimi K3, through an API.
2. You want a low cost per task: DeepSeek V4 Pro, through its API or OpenRouter.
3. You want everything local, on your own GPU: Qwen 3.8 27B, knowing that on long agentic tasks it stays far from proprietary models.

Benchmarks only tell part of the story. Before switching models, test them on the tasks you actually do every day: that’s where you find out which tier you need.