OpenClaw 2026.4.23 Release Notes — Kimi K2.6 / Qwen 3.6-27B / GPT-5.5 Integrations and Agent Improvements

Between April 16 and April 24, 2026, the AI industry shipped Claude Opus 4.7 (4/16), Kimi K2.6 GA (4/21), Gemini 3.1 Pro Deep Research / Deep Research Max (4/21), Qwen 3.6-27B (4/22), GPT-5.5 (4/23), DeepSeek V4 Preview (4/24), and the open-sourcing of Japan's government AI "Genai" (4/24) — all in one week. OpenClaw 2026.4.23 lands in that wave, expanding the agent platform on Mac mini to integrate with these new models for both local and cloud-hybrid workflows.

OpenClaw can now drive a wider matrix of models: - Open-weight, local on Mac mini: Kimi K2.6 (quantized), Qwen 3.6-27B Dense, full Gemma 4 family, Qwen 3.5-9B, DeepSeek V4-Flash (lightweight) - Cloud APIs: Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro - Government / public sector: connector templates for Japan's open-sourced "Genai" (genai-web / genai-ai-api) Clients can mix local and cloud freely based on confidentiality and cost needs.

Following the April 21 launch of Google's Gemini 3.1 Pro Deep Research and Deep Research Max APIs, OpenClaw ships a new "research agent" preset. It combines internal RAG and web search through agent swarms for long-horizon research tasks, and pairs naturally with Kimi K2.6's swarm coordination to run cross-source investigations reliably.

The hidden theme of this release is cost. We strengthened the model router so OpenClaw can dispatch lightweight tasks (e.g., classification) to local Qwen 3.5-9B / DeepSeek V4-Flash and only escalate to Claude Opus 4.7 or GPT-5.5 when design-level judgment is required. In our internal measurements, this can compress monthly cloud API spend by up to 80% (workload-dependent).

Fully on-premise mode remains the default for sensitive workloads — local-mode operation does not send data outside your network. Genai (`genai-web`) connector templates are intended for government, local-government, and quasi-public sector pilots. Existing sandboxing, audit logs, and permission controls are unchanged.

Existing customers can update from the admin console. New customers can start with a Mac mini M4 16GB+ entry configuration. See OpenClaw service page or Contact for details.

Q1: Can Kimi K2.6 and Qwen 3.6-27B run on a Mac mini? A: With Q4 quantization, an M4 Pro 32GB / M3 Max 64GB-class machine handles them in production. Full-precision needs a GPU server. Q2: Isn't mixing cloud APIs and local models complex to manage? A: OpenClaw's router dispatches per-task automatically. Manual override is also available — all from the Web UI. Q3: What does the Genai integration actually do? A: We bundle templates that let the genai-web UI call OpenClaw's agent APIs. Works across AWS, Azure, and Google Cloud. Q4: Will the update break existing workflows? A: Backward compatibility is preserved — existing agent definitions continue to work.

- Claude Opus 4.7 (Anthropic, 2026/4/16) - Kimi K2.6 GA (Moonshot AI, 2026/4/21) - Deep Research / Deep Research Max (Google, 2026/4/21) - Qwen 3.6-27B (Alibaba, 2026/4/22) - GPT-5.5 (OpenAI, 2026/4/23) - DeepSeek V4 Preview (DeepSeek, 2026/4/24) - Government AI "Genai" open-sourced (Digital Agency, 2026/4/24)