← All news·2026-06-21·4 min read

Free 1 Trillion Parameter Model That Writes Code Like A Pro

Kimi K2.7 Code is a 1T parameter model that runs locally, scores 62 on a coding test, and does so while thinking ~30% less.

open weightscoding modelsMoonshot AIKimiAI agents

Moonshot AI’s latest model, Kimi K2.7 Code, is an open-weights model focused on coding and coding agents. Despite being a massive model with 1 trillion parameters, it only uses 32 billion of those parameters per request, allowing it to achieve its goal of “big brain, small bill”. This means that despite being a huge model, it’s able to run quickly because it’s only using the necessary portion of the model at a given time.

Kimi K2.7 Code represents a significant improvement over the original Kimi model. For example, on Moonshot’s Kimi Code Bench v2, it achieves a score of 62.0, which is a substantial improvement over the original model’s score of 50.9. Additionally, it improves its performance on various tool-use benchmarks, achieving a score of 76.0 compared to the original model’s 69.4, and a score of 81.1 compared to the original model’s 72.8. It’s also able to maintain approximately 256,000 tokens of context, allowing it to remember a sizable portion of a code base, and it reduces its ‘thinking’ tokens by ~30% compared to the previous model, allowing it to find solutions more quickly and improve overall performance on tasks.

Open weights are the real story

One of the major advantages of this model is that it’s open-weights, meaning that the model weights are publicly available under a Modified MIT License. Having access to the model weights allows you to deploy the model locally, without having to send your code to a third party, and without having to pay for each request. This model works with popular open inference engines such as vLLM, SGLang, and KTransformers, and it also comes with a small vision model, allowing it to process images and screenshots alongside text.

One of the main benefits of this model for developers is that it works with existing agent tooling that many people may already be using, such as the MCP protocol, which allows an AI to interact with external tools like browsers, files, and databases. The significant improvements on MCP benchmarks indicate that this model is better suited for agent tasks that require multiple steps and interactions, rather than just generating individual functions.

Why it matters for you

In the past, the only way to get high-quality coding assistance was to pay a premium for access to a few closed services and pay for each request. With this free, downloadable model that performs well and allows you to run your code locally, we’re starting to see a shift away from these walled gardens. Now, if you’re someone who codes with a coding agent, you have an option that is both powerful and free from vendor lock-in and usage restrictions.

The reduction of thinking tokens by ~30% is the hidden gem here. In many cases, reasoning models perform well on benchmarks by “thinking” a lot, which results in slower and more expensive performance in practice. By achieving better results with less thinking, we’re able to make a model that is enjoyable to use.

ℹWhat this means in practice

'Open weights' is the phrase to remember: it means you can download the model and run it on your own machine, keeping your code private and skipping per-request fees. Kimi K2.7 Code is a near-top-tier coding model you can self-host — worth a look if you're tired of metered, closed assistants.

Source: huggingface.co

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids