Agents
Updated

Best uncensored AI for coding in 2026

The best uncensored AI for coding is Kimi K3.1 Pro, an abliterated 2,780B-parameter model with a 1M-token context that runs in Claude Code, Codex, OpenCode and Cursor.

8 models6 coding agents7 APIs

At a glance

Four hosted models, then four you run yourself.

How we rank
#ModelParametersContextMax outputPrice per 1M, in / outInputsPick
Hosted, through an API
1Kimi K3.1 ProMoonshot AI · abliterated2,780B1M943,718$3.00 / $8.00Text, Images, VideoBest overall
2DeepSeek V4.1 Flash 1017DeepSeek · abliterated763B1M131,072$0.24 / $0.90Text, ImagesBest value
3GLM-5.3Z.ai · abliterated753B1M131,072$2.00 / $4.00TextBest text-only coder
4GLM-5.3 FlashZ.ai · abliterated320B total, 18B active1M131,072$0.24 / $1.00Text, ImagesBest for volume
Local, on a GPU you own
5Qwen 3.8 27B Abliteratedhuihui-ai · runs locally27.8BYour memoryYour serverabout 17 GB at 4-bitTextBest local overall
6Qwen 3.6 35B-A3B Abliteratedhuihui-ai · runs locally36B total, 3B activeYour memoryYour serverabout 22 GB at 4-bitTextBest local mixture-of-experts
7Gemma 4 31B Hereticcoder3101 · runs locally31.3BYour memoryYour serverabout 19 GB at 4-bitTextBest local Gemma
8Llama 3.1 8B Abliteratedmlabonne · runs locally8BYour memoryYour serverabout 5 GB at 4-bitTextBest for a laptop

Hosted prices are Silk Compute's, read November 18, 2026, and 4-bit sizes are estimates.

Why each model ranks where it does

05Best local overall

Qwen 3.8 27B Abliterated

huihui-ai/Huihui-Qwen3.8-27B-abliterated

The newest abliterated Qwen here, small enough at 4-bit to leave room for context on a 24 GB card.

Hardwareone 24 GB GPU or a 32 GB Mac
LicenseApache 2.0
07Best local Gemma

Gemma 4 31B Heretic

coder3101/gemma-4-31B-it-heretic

Google's Gemma 4 31B, edited to stay close to the original model.

Hardwareone 24 GB GPU
LicenseApache 2.0
08Best for a laptop

Llama 3.1 8B Abliterated

mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated

The smallest and oldest model here, so expect much less from it in an agent than from the 27B to 36B models.

Hardwarean 8 GB GPU or a laptop
LicenseLlama 3.1 Community License

Set it up in one file

Each config adds every hosted model and starts on DeepSeek V4.1 Flash 1017.

Step-by-step guides
~/.claude/settings.json
{
  "apiKeyHelper": "echo $SILK_API_KEY",
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.silkcompute.com",
    "ANTHROPIC_DEFAULT_MODEL": "deepseek-v4.1-flash-1017-uncensored",
    "ANTHROPIC_DEFAULT_FABLE_MODEL": "kimi-k3.1-pro-uncensored",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "kimi-k3.1-pro-uncensored",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "deepseek-v4.1-flash-1017-uncensored",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "deepseek-v4.1-flash-1017-uncensored",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "1048576"
  },
  "modelPicker": {
    "replaceBuiltInOptions": true,
    "options": [
      { "model": "kimi-k3.1-pro-uncensored", "label": "Kimi K3.1 Pro" },
      {
        "model": "deepseek-v4.1-flash-1017-uncensored",
        "label": "DeepSeek V4.1 Flash 1017"
      },
      {
        "model": "deepseek-v4.1-flash-1017-fast-uncensored",
        "label": "DeepSeek V4.1 Flash 1017 Fast"
      },
      { "model": "glm-5.3-uncensored", "label": "GLM-5.3" },
      { "model": "glm-5.3-flash-uncensored", "label": "GLM-5.3 Flash" },
      { "model": "hy4-uncensored", "label": "Hy4" }
    ]
  }
}
~/.codex/config.toml
model = "deepseek-v4.1-flash-1017-uncensored"
model_provider = "silkcompute"
web_search = "disabled"

[model_providers.silkcompute]
name = "Silk Compute"
base_url = "https://api.silkcompute.com"
env_key = "SILK_API_KEY"
model_catalog_url = "https://api.silkcompute.com/codex/models"
opencode.json
{
  "$schema": "https://opencode.ai/config.json",
  "model": "silkcompute/deepseek-v4.1-flash-1017-uncensored",
  "provider": {
    "silkcompute": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Silk Compute",
      "options": {
        "baseURL": "https://api.silkcompute.com",
        "apiKey": "{env:SILK_API_KEY}"
      },
      "models": {
        "kimi-k3.1-pro-uncensored": {
          "name": "Kimi K3.1 Pro",
          "tool_call": true,
          "interleaved": { "field": "reasoning_content" },
          "modalities": { "input": ["text", "image", "video"], "output": ["text"] },
          "limit": { "context": 1048576, "output": 943718 },
          "cost": { "input": 3, "output": 8, "cache_read": 0.3 },
          "variants": {
            "none": { "reasoningEffort": "none" },
            "low": { "reasoningEffort": "low" },
            "high": { "reasoningEffort": "high" },
            "max": { "reasoningEffort": "max" }
          }
        },
        "deepseek-v4.1-flash-1017-uncensored": {
          "name": "DeepSeek V4.1 Flash 1017",
          "tool_call": true,
          "interleaved": { "field": "reasoning_content" },
          "modalities": { "input": ["text", "image"], "output": ["text"] },
          "limit": { "context": 1048576, "output": 131072 },
          "cost": { "input": 0.24, "output": 0.9, "cache_read": 0.038 },
          "variants": {
            "none": { "reasoningEffort": "none" },
            "low": { "reasoningEffort": "low" },
            "high": { "reasoningEffort": "high" },
            "max": { "reasoningEffort": "max" }
          }
        },
        "deepseek-v4.1-flash-1017-fast-uncensored": {
          "name": "DeepSeek V4.1 Flash 1017 Fast",
          "tool_call": true,
          "interleaved": { "field": "reasoning_content" },
          "modalities": { "input": ["text", "image"], "output": ["text"] },
          "limit": { "context": 1048576, "output": 131072 },
          "cost": { "input": 0.24, "output": 0.9, "cache_read": 0.05 },
          "variants": {
            "none": { "reasoningEffort": "none" },
            "low": { "reasoningEffort": "low" },
            "high": { "reasoningEffort": "high" },
            "max": { "reasoningEffort": "max" }
          }
        },
        "glm-5.3-uncensored": {
          "name": "GLM-5.3",
          "tool_call": true,
          "interleaved": { "field": "reasoning_content" },
          "modalities": { "input": ["text"], "output": ["text"] },
          "limit": { "context": 1048576, "output": 131072 },
          "cost": { "input": 2, "output": 4, "cache_read": 0.3 },
          "variants": {
            "none": { "reasoningEffort": "none" },
            "low": { "reasoningEffort": "low" },
            "high": { "reasoningEffort": "high" },
            "max": { "reasoningEffort": "max" }
          }
        },
        "glm-5.3-flash-uncensored": {
          "name": "GLM-5.3 Flash",
          "tool_call": true,
          "interleaved": { "field": "reasoning_content" },
          "modalities": { "input": ["text", "image"], "output": ["text"] },
          "limit": { "context": 1048576, "output": 131072 },
          "cost": { "input": 0.24, "output": 1, "cache_read": 0.06 },
          "variants": {
            "none": { "reasoningEffort": "none" },
            "low": { "reasoningEffort": "low" },
            "high": { "reasoningEffort": "high" },
            "max": { "reasoningEffort": "max" }
          }
        },
        "hy4-uncensored": {
          "name": "Hy4",
          "tool_call": true,
          "interleaved": { "field": "reasoning_content" },
          "modalities": { "input": ["text", "image"], "output": ["text"] },
          "limit": { "context": 1048576, "output": 65536 },
          "cost": { "input": 1.6, "output": 3.2, "cache_read": 0.3 },
          "variants": {
            "none": { "reasoningEffort": "none" },
            "low": { "reasoningEffort": "low" },
            "high": { "reasoningEffort": "high" },
            "max": { "reasoningEffort": "max" }
          }
        }
      }
    }
  }
}

The configs read your key from SILK_API_KEY and also add Hy4, which is not ranked here.

APIs for coding agents compares where to run them.

Questions

Can I use an uncensored model in Claude Code?

Yes. Claude Code reads ANTHROPIC_BASE_URL, so any API that serves an abliterated model in the Anthropic Messages format works, as the Claude Code setup page shows.

Is an abliterated model worse at coding?

A little, but the size of the base model matters far more, as abliterated vs fine-tuned explains.

What is an uncensored coding model for?

Legitimate work that hosted assistants often decline, such as security tooling for systems you are authorized to test, reverse engineering your own software and scrapers for public data. It is not a way around the law or anyone else's permission.