Best Uncensored Models for Coding Agents: How to Pick One

Why does a refusal hurt more in an agent?

A refusal in a chat costs you one message. A refusal in an agent can cost you the run. You start a long job, go make dinner, and come back to find it stopped at step 40 with a polite note about what it can't help with.

It happens on ordinary work, too: a security fix, a fuzzer, a test fixture with a real exploit string in it, a scraper. Mainstream models are trained to decline broadly, and an agent reads whatever the task touches. Sooner or later, it reads something that trips the wire.

What does a coding agent need?

Abliteration removes the no, and that's only the start. If you're choosing a model for an agent, look for these four things:

• Tool calling. An agent is stuck without it. Every model we list supports tool calling and streaming.
• Context. Agents read a lot. Every Silk Compute model has 1M tokens of context, so most of a repo stays in view.
• A low cached input price. Agents resend their whole context every turn, so this is the number your bill really rides on.
• The right API format. Each agent speaks its own. Silk Compute speaks all three.

Which uncensored model should you use for which job?

Here's the short version. The first five are on Silk Compute today, and the last two run on your own machine.

Kimi K3.1 Pro takes the hardest workloads, with up to 944K tokens in one reply. Our Hy4 launch post scores Hy4 against it. GLM-5.5 is coming soon. For more on the two we'd start with, see DeepSeek V4.1 Flash 1017 and GLM-5.3 Flash.

Each maker's own table has a row worth reading. DeepSeek reports 77.0 on DeepSWE v1.1, a coding benchmark, the top score on that row of its own table. Z.ai reports 84.5 for GLM-5.3 on CyberGym, a vulnerability discovery benchmark, the top score on that row of its own table. Both are for the original models, before abliteration, and abliteration can shift things a little.

What does an uncensored coding agent cost?

Less than the sticker price suggests. The bars show 1M tokens on each of our models today, at the mix on our pricing page: 97% cached input, 2% uncached input, 1% output. That mix is an assumption, and your own will vary.

So don't shop on input price. An agent barely pays it. GLM-5.3 Flash has the lowest input price of the five, and DeepSeek still has the lower bill, because its cached input costs less.

If you'd rather have a set monthly spend, plans start at $10 a month and cover more usage than they cost.

Can you run an uncensored coding agent at home?

Yes, if the model fits. The trouble is that the big ones don't. DeepSeek V4.1 Flash 1017 is a 501 GB file at 4-bit, and GLM-5.3 is 467 GB. Our abliterated AI size chart shows what fits on a laptop, a gaming GPU or a big workstation. You'll also need something local that speaks the API format your agent expects.

How do you set it up?

Each agent needs one config file, and each guide has it ready to paste. Set your key, pick a model, and go. Your account runs up to 8 requests at once, so subagents work in parallel.

• Claude CodeSwitch with /model and /effort
• CodexSwitch the model and effort with /model
• OpenCodeSwitch with /models, effort with ctrl+t
• Other toolsKilo Code, Cline, Roo Code and more

Tips for long agent runs

• Match the effort to the step. High is the default. Low keeps routine steps quick, and max is for the hard ones.
• Try a small job first. An abliterated model can still decline, and it's a different model, not the same one minus the no. Give it a contained task before you hand it the whole repo.
• Watch the right bill. Claude Code's own cost readout assumes Claude's prices, so it overstates the bill on these models. Check the Usage page in your account instead.
• Plan around server tools. Anthropic server tools such as web search aren't available here, so wire in your own if the agent needs them.