Unrestricted
AI agent
To build an unrestricted AI agent, point an open-source agent framework at an abliterated model through its base URL. Use Kimi K3.1 Pro for long runs, or DeepSeek V4.1 Flash 1017 at $0.24 per million input tokens for value.
- Kimi K3.1 Pro for long runs
- DeepSeek V4.1 Flash for value
- 11 frameworks, sources linked
Pick the model by how long the agent runs
All models and run costsKimi K3.1 Pro
The largest model here at 2,780B parameters, for runs that read and write a lot.
- Context
- 1,048,576
- Max output
- 943,718
- Input
- Text, images, video
- Per 1M in / out
- $3.50 / $8.00
DeepSeek V4.1 Flash 1017
763B parameters at $0.24 per million input tokens, for steps that repeat thousands of times.
- Context
- 1,048,576
- Max output
- 131,072
- Input
- Text, images
- Per 1M in / out
- $0.24 / $0.90
| Also served | Input | Per 1M in / out | Note |
|---|---|---|---|
| GLM-5.3 | Text | $2.00 / $4.00 | Text only, strong at software work. |
| GLM-5.3 Flash | Text, images | $0.24 / $1.00 | 320B total with 18B active, for volume. |
Pick the framework
All 11 setupsOpenAI Agents SDK
Agents, handoffs and sessions in Python, on any Chat Completions endpoint.
smolagents
Small agents that act in code or JSON tool calls, with a browser example.
browser-use
A browser agent that writes Python against Chrome, with any compatible model.
Microsoft Agent Framework
The successor to AutoGen, with a Chat Completions client that takes a base URL.
Letta
Stateful agents with a memory filesystem, on any endpoint with tool calling.
What each framework supports
| Framework | OpenAI format | Anthropic format | Tool calling | Streaming | Browser use | Memory |
|---|---|---|---|---|---|---|
| smolagentsPython SDK | ||||||
| browser-useBrowser agent | ||||||
| Microsoft Agent FrameworkPython SDK | ||||||
| LettaApp |
When the agent still stops
Why agents stop on refusals
A refusal returns text with no tool call, so the loop ends.
GuideTool calling with abliterated models
Both formats, side by side, with a request in each.
GuideRun agents locally
Four abliterated open-weight models and the base URL for each framework.
RankingBest API for agents
Seven uncensored APIs, ranked on what an agent loop needs.