3 New Open AI Models You Can Run Yourself (October 2026)

Three open models released between September 21 and October 8, 2026 cover coding, images and general chat: JetBrains Mellum2.1, Alibaba's Qwen-Image-2.1 and Mistral Large 4. The first two are small enough for your own machine; Mistral Large 4 is open but sized for servers.

The three models at a glance

• Mellum2.1 (JetBrains, October 8): 12B mixture-of-experts with 2.5B active parameters, Apache 2.0 licence, on Hugging Face. For coding agents and fast sub-agents.
• Qwen-Image-2.1 (Alibaba, September 21): open source, with a 7B image-generation component. For generating and editing images, including transparent PNGs.
• Mistral Large 4 (Mistral AI, October 6): open weights, 1.05 trillion parameters in total with 52B active, plus a 1.6B vision encoder and a 1-million-token context. Still a public preview.

Mellum2.1: a coding model for your own hardware

JetBrains trained Mellum2.1 mostly with reinforcement learning in real coding sandboxes, so it can explore a repository, edit files and check its own changes. It keeps Mellum2's fast architecture, and JetBrains says multi-token prediction makes a single request about 1.6 times faster. Use it as a cheap worker next to a bigger model, or on its own when your code must not leave your machine.

Qwen-Image-2.1: transparent images and 10 references

Qwen-Image-2.1 generates and edits images in one model. It can output an image with a transparent background straight from a prompt, cut a subject out of a photo, and combine up to 10 reference images — handy for virtual try-on or furnishing a room from product shots. You can point an edit at one region with a circle, a painted mark or a separate mask.

Mistral Large 4: open, but not for a laptop

Mistral Large 4 is a multimodal model with open weights. With over a trillion parameters in total it needs server-class hardware, so most people will use it through Mistral's API, currently on sale at $0.68 per million input tokens and $2.09 per million output tokens.

How to start with each model

1. Mellum2.1: download it from the JetBrains collection on Hugging Face and serve it with vLLM. GGUF builds for llama.cpp, Ollama and LM Studio are announced as coming soon.
2. Qwen-Image-2.1: start from the official announcement, which links the weights, and test a transparent logo or a product cut-out first.
3. Mistral Large 4: try it in the Mistral playground under the name mistral-large-4 before deciding to host it yourself.
4. No GPU? WorkCrafter's code tool and image tool run hosted models in the browser, so you can compare results before buying hardware.

Our tips for open models

• Pick by task, not by size: a coding model with 2.5B active parameters can beat a giant on speed and cost for routine fixes.
• Read the licence before shipping a product; Apache 2.0, used by Mellum2.1, allows commercial use.
• Preview models change: pin the exact version you tested.
• Keep a hosted fallback for the days your local machine is busy or down.

More on open models and WorkCrafter

• The AI Code Tool: Write and Debug Faster
• Why Free AI Tools Aren't Really Free
• AI Content Creation Tools: A Comprehensive Guide for 2026
• Open the code generator

Sources on open models

• Mellum2.1 Gets to Work: A Fast Open Model for Coding Agents — JetBrains, October 2026. Size, licence, training, speed figure, GGUF status.
• Qwen-Image-2.1: Compact, Efficient, and Unified Image Creation — Alibaba Cloud, September 21, 2026. 7B component, transparency, 10 references, mask editing.
• Mistral Large 4 — Mistral AI documentation, October 6, 2026. Parameters, vision encoder, 1M context, public preview, API prices.