Models
Autohand models
Fantail and Moa are Autohand's first-party coding models. Use Fantail when latency matters. Use Moa when the task needs more context, planning, or repository-level reasoning. For hosted coding agents, Autohand also recommends a curated set of high-context open and partner models.
Choose a model
| Model | Use it for | Context | Serving profile |
|---|---|---|---|
| Fantail | Completions, quick fixes, review comments, test scaffolds, and short agent loops. | 32K | Latency-first, low-cost coding assistance. |
| Moa | Large refactors, architecture work, security reviews, and multi-file planning. | 200K | Reasoning-first execution for complex software engineering tasks. |
Recommended coding models
Use these models for coding agents, long-horizon repo work, tool-heavy sessions, and hosted team workflows.
| Model | Provider | Best for | Context | License | Top signal |
|---|---|---|---|---|---|
| Autohand Fantail Pro | Autohand | Fast coding loops, completions, reviews, and short agent tasks. | 32K | Hosted | Latency-first Autohand model |
| Autohand Moa | Autohand | Larger refactors, planning, and multi-step repository changes. | 200K | Hosted | Reasoning-first Autohand model |
| GLM-5.1 | Z.ai | Long-horizon agentic engineering. | 200K | MIT | Terminal-Bench 2.0 SOTA |
| GLM-5.2 | Z.ai | 1M-context long-horizon agentic engineering. | 1M | MIT | Terminal-Bench 2.1 81.0 |
| Sakana AI Fugu | Sakana AI | Balanced low-latency multi-agent coding and chat. | 1M | Orchestrated | SWE-Bench Pro 59.0 |
| Sakana AI Fugu Ultra | Sakana AI | Multi-agent orchestration across frontier models. | 1M | Orchestrated | SWE-Bench Pro 73.7 |
| Kimi K2.6 | Moonshot AI | Agent swarms and long autonomous runs. | 256K | Modified MIT | SWE-Bench Pro 58.6% |
| DeepSeek V4 Pro | DeepSeek | 1M-context coding and competitive programming. | 1M | MIT | LiveCodeBench 93.5 |
| Qwen3-Coder-Next | Alibaba Qwen | Efficiency per active parameter. | 256K | Apache 2.0 | SWE-Bench Verified 71.3 |
| Qwen3.6-27B | Alibaba Qwen | Dense repo-level coding. | 262K | Apache 2.0 | SWE-Bench Verified 77.2 |
| MiniMax M2.5 | MiniMax | Free hosted open-weight coding. | 200K | Open weights | Free hosted variant |
| GLM-5 | Z.ai | Local self-hosting foundations. | 200K | MIT | #1 Vending Bench 2 OSS |
| Devstral 2 | Mistral AI | Mistral coding workloads. | 256K | Apache 2.0 | SWE-Bench Verified 72.2 |
| Devstral Small 2 | Mistral AI | Consumer GPU coding workflows. | 128K | Apache 2.0 | SWE-Bench Verified 68% |
| Trinity Large Thinking | Arcee AI | US-origin open reasoning and long agent loops. | 128K | Apache 2.0 | #2 PinchBench |
Use a model
Pass the model identifier where you start the run. The identifiers are fantail and moa.
Fantail
Fantail is built for interactive coding where feedback speed matters. It is the best default for small, local edits and repeated assistant loops.
- Best for autocomplete, quick bug fixes, code review notes, documentation snippets, and small tests.
- Works well when the relevant files are already in the prompt or easy for the agent to inspect.
- Use Moa instead when the task depends on broad architecture or a large amount of repository context.
Moa
Moa is built for deeper software engineering work. It has a larger context window and is a better fit when the model needs to reason across many files or constraints.
- Best for large refactors, migration plans, security audits, technical design, and repository analysis.
- Use it when the model needs to compare patterns across a codebase before making changes.
- Switch back to Fantail for final polish, small follow-up edits, and fast review cycles.
Common patterns
- Start fast. Use Fantail for the first pass when the task is scoped to a file, function, or failing test.
- Escalate for context. Switch to Moa when the answer depends on architecture, cross-package contracts, or hidden coupling.
- Keep the model explicit. Include the model in scripts and SDK calls so runs are repeatable.