π― Model Selection
π οΈ Deployment Guide: Which Model to Pick? (Model Landscape)
OpenClaw is model-agnostic, but the community's "heart" is quite vocal. Here is the deployment landscape for early 2026:
1 π₯ Claude (Anthropic) β The "Soulmate": - Popular Choice: Claude Sonnet 4 / 4.5. - Why: Widely considered the most stable for tool-calling logic and following Lobster-specific instructions. Despite the recent legal drama, its coding capability remains the heart of choice for most.
2 π₯ GPT (OpenAI) β The "Veteran": - Popular Choice: GPT-4o / GPT-5.4 (Thinking). - Why: Even with Peter moving to OpenAI, GPT's rigor in code generation and community support remain top-tier. The o1-series' reasoning shines in complex automation tasks.
3 π₯ Gemini (Google) β The "King of Context & Value": - Popular Choice: Gemini 2.5 / 3 Flash / Pro. - Why: With 1M+ long context, it has an unmatched advantage when dealing with massive log files. Gemini 3 Flash hit a staggering 95.1% success rate in recent OpenClaw task benchmarks.
4 π DeepSeek / Kimi / MiniMax β The "Rising Stars": - Why: Extremely fast deployment in the APAC region and aggressive pricing (often with free tier events). The savior of low-budget "Cloud Lobster Farming."
5 π Ollama (Local) β The "Privacy Guard": - Why: The latest version now supports MiniMax, Kimi, GLM, and Qwen3. Perfect for privacy-conscious developers wanting to explore the latest model ecosystem locally.
π§ Field Notes: Model Parameters vs. "Lobster Power"
π₯οΈ Official NVIDIA Recommendations: DGX Spark Local Setup Guide
The official OpenClaw on DGX Spark instructions provide the following suggested models:
| GPU memory | Suggested model | Model size | Notes |
|---|---|---|---|
| 8β12 GB | qwen3-4B-Thinking-2507 |
~5GB | β |
| 16 GB | gpt-oss-20b |
~12GB | Lower latency, good for interactive use |
| 24β48 GB | Nemotron-3-Nano-30B-A3B |
~20GB | β |
| 128 GB | gpt-oss-120b |
~65GB | Best quality on DGX Spark (quantized); leaves ~63GB for context window and other processes; use 20B/30B if you prefer faster responses |
Quality vs. latency: The 120B model gives the best accuracy and capability but has higher per-token latency. If you prefer snappier replies, use
gpt-oss-20b(or a 30B model) instead; both run comfortably on DGX Spark with plenty of memory headroom.
General Community Parameter Experience
| Parameters | Tier | Field Notes (Lobster Power) |
|---|---|---|
| < 30B (e.g., 9B, 14B, 27B) | π¬ Chatty | Great for chatting, but fails to meet human "Agent" expectations. Tool-calling logic is often broken. |
| 30B - 70B | π οΈ Assistant | Basic operations OK. Competent for searches, reminders, and simple tasks. The "sweet spot" for mid-range local. |
| 70B - 100B | πΌ Professional | Handles complex workflows. Good for Office automation and business processes. Can write simple code, but struggles with large Skills. |
| > 100B+ (Cloud SOTA) | π¦ The Ultimate King | True Agentic Freedom. Multimodal; handles coding/PRs/iteration solo. The only tier for "Set it and forget it" autonomy. |
Lobster Insight: Local is for privacy and thrift; Cloud is for true "Agentic Freedom." Unless you have 128GB+ RAM at home, leave the heavy lifting to the cloud brains. π¦π‘
Quick Tip:
-
For beginners, start with Claude Sonnet for maximum stability. When handling massive project repos, switching to Gemini Pro will feel like a different dimension.
-
β οΈ Security Warning: If you're using cloud models, guard your API Keys with your life! The recent $82,000 hijacking incident is a brutal wake-up call (see Latest Crisis); always set strict usage quotas.
-
Local Players: The latest Ollama now supports the MiniMax, Kimi, GLM, and Qwen3 families. This is ideal for users with Mac mini (64GB+ RAM suggested) or DGX Spark (128GB), allowing you to forget about API bills entirely while ensuring maximum privacy. Even if you don't run them locally, their cloud APIs are extremely budget-friendly. It's the best option for Lobster Farmers to avoid bill shock. π¦β¨
-
Context Size: Regardless of the model, it is recommended to set the Context Size to at least 64K to ensure the Lobster maintains its "memory" during long sessions. π¦π§