Models we shipand why.
Zaguán Blade routes your requests through the Zaguán Coder Daemon, which supports a curated set of hosted frontier models. Each model is tested for coding, tool use, and stability before it is enabled.
Updated 2026-07-07
Enabled Models
Only models marked enabled: true in the production registry appear in the model picker. The list below is the current supported set.
Free: Tencent Hy3
Strong coding model with balanced performance.
Currently offered free of charge. A strong coding model with a balanced profile — a good way to try heavier tasks in Zaguán Blade at no cost.
MiMo v2.5 Pro
Flagship agentic model.
This is a very good model. It's fast and capable on almost every task. One of it's stongest points at the moment is its price.
MiMo v2.5
Great agentic model.
This is a very good model for its size. It's fast and capable on almost every task. One of it's stongest points at the moment is its price. I use this model for smaller tasks due to its speed.
Cohere Command A+
Strong MoE coding model.
This is a very good model from Cohere. It's decent at webdesign. Not excellent. I use this for small to medium tasks and trust it to do these tasks.
DeepSeek v4 Pro
Large reasoning model.
This is a really good model that I've used for a few heavy tasks. It's very performant for its size and it's webdesign skills are above average. It's a reasoning model, so it spends a lot of time thinking. However, since the price is so low it really is worth it.
DeepSeek v4 Flash
Fast reasoning model.
This model is a lot faster than its big brother. It's also very capable and I use it mostly for small to medium tasks due to its speed and accuracy. Its webdesign skills are above average.
Mistral Large 3
Mistral agentic coding model with strong tool use.
This is a very capable model. It's a bit slower than many of the other models. I use it sometimes when I want to experiment with things. It's not one of my go-to models. This model is about average on webdesign.
Mistral 3.5 Medium
Mistral agentic coding model with strong tool use.
This is a much faster model than its brother. I turn to this more that Mistral 3. Its webdesign skills is above average.
Claude Sonnet 5
Balanced performance and speed.
The latest Sonnet from Anthropic. Balanced speed and quality for everyday coding tasks, with strong tool use and reliable output.
Claude Opus 4.8
Latest model from Anthropic.
This is the workhorse. It's not the fastest, but it is very stable and always get the job done and done well. Its webdesign skills are very Claude. Adequate.
Claude Fable 5
Fable 5 from Anthropic.
Anthropic's latest model with a massive context window and reasoning capabilities. A strong alternative to Opus for heavy work.
Qwen 3.7 Max
Strong coding performance.
This is a big model and it shows. It's a bit slow. However, it does the tasks I've asked it to do. Its webdesign skills are slightly above average.
Qwen3.7 Plus
Strong coding performance.
This is a great model. It's fast. It's capable and it gets the job done. I tend to turn to this model more than Max. Its webdesign skills are about average.
GPT 5.5 (medium)
OpenAI coding model with medium reasoning.
This is, by far, the most popular model for high quality code and output and one model I've turned to many times. It's webdesign skills is close to what we expect as AI-slop, but with the right harness and great skills it's easy to guide it to do better.
Kimi K2.7 Code
Kimi K2.7 Code.
This is one of my go-to models for many tasks. It's very capable and very fast. It's webdesign skills are about average and with the correct harness and skills goes beyond. Kimi claims that is uses about 30% less tokens than previous models.
GLM-5.2 (max)
Strong coding model — max reasoning effort.
The max reasoning variant of GLM-5.2. It spends more time thinking through complex problems. Through this provider it's still fast despite the heavy reasoning. Best for tasks that benefit from deep analysis before acting.
GLM-5.2 (high)
Strong coding model — high reasoning effort.
This model is huge, however through the provider we're using it's very fast and very capable. This is one of the models I turn to when I need to do heavy lifting. Its webdsign skill is about average and with the correct harness and skills becomes much better. This model should not be used for small tasks that requires speed and fast changes. It is meant for long running and complex tasks.
GLM-5.2 Instruct
Strong coding model — instruction mode.
GLM-5.2 in pure instruction-following mode without reasoning overhead. Fast and direct, ideal when you want straightforward answers without the model thinking out loud.
MiniMax M3
Strong coding model with balanced performance.
This is another huge model that I often turn to for heavy lifting. It's fast and very capable. Its webdesign skill is above average and with the correct harness and skills becomes very good.
Not sure what to choose?
These are sane defaults drawn from the registry above. Your own workload, token budget, and tolerance for reasoning time may change the answer.
GPT 5.5 (medium)
The most popular model for high-quality code and output; a safe first choice for most tasks.
MiMo v2.5
Very capable for its size and one of the best price-to-performance ratios in the registry.
GLM-5.2
Huge context window and fast through the provider; good for heavy lifting and large files.
Kimi K2.7 Code
Very capable and very fast — a go-to model for many tasks. Uses about 30% fewer tokens than previous generations.
DeepSeek v4 Pro
A large reasoning model that performs well for heavy tasks at a low price.
Ollama & OpenAI-compatible servers
Connect your own local models through Ollama or any OpenAI-compatible server. Read the docs to set it up.
Why a model might not be here
The registry is intentionally small. Models are excluded when they fail our criteria, not because we forgot about them.
Instability
Unreliable API behavior, frequent provider changes, or broken streaming make a model hard to ship.
Weak coding or tool use
Poor code generation, bad tool calling, or inconsistent reasoning are deal-breakers for a coding assistant.
Cost without quality
High price only makes sense if the output quality justifies it. We drop models that cost more but deliver less.
Context-window worship
Large context windows help, but they do not replace good retrieval, skills, and disciplined context selection.
What the fields mean
Provider
The backend provider that serves the model. Zaguán Coder Daemon handles provider-specific routing, prompt templates, and reasoning formatting for you.
API Format
The protocol the provider speaks: chat_completions,responses, or messages. Blade abstracts this so the UI does not need to know the difference.
Context Window
The maximum number of tokens the model can accept in a single request. Blade tries to keep context small and targeted, so large windows are used for headroom, not bloat.
Capabilities
What the model is currently configured to do. Tool use means the model can invoke Blade's code-intelligence and editing tools. Reasoning means it emits intermediate thinking.