Introducing the First Public Beta Release for Intel SuperClaw
In case you missed it, Intel SuperClaw recently debuted with its first public beta release: https://aibuilder.intel.com/#/superclaw!
Computex 2026 marked a turning point for Intel SuperClaw. On the show floor, customers weren't just talking about hybrid agentic AI - they were running real-life agentic AI use cases, powered by Intel SuperClaw, on both AI PCs and edge servers at their booths. With the release of the SuperClaw v1.0 beta, this momentum is culminating in a significant milestone: partners and customers running real agentic workloads on Intel platforms, validating the innovative work and dedication the Intel AI solution team has put into SuperClaw to-date.
The initial beta release for Intel SuperClaw is designed for an Intel workstation/edge server setup utilizing two Intel systems:
- Intel AI PC client system, where the SuperClaw desktop app runs.
- Minimum recommended spec is an Intel® Core™ Series 3-powered device with 16GB RAM.
- Intel model-serving workstation system, which will run the Qwen3-Coder-Next-80B model.
- Minimum recommended spec is an Intel workstation running four (4) Intel® Arc™ Pro B70 GPUs.
With this setup, the SuperClaw UI will be running on the AI PC while the 80B model workload is running on the B70-based workstation or edge server system - a true device-to-edge agentic pipeline, entirely on Intel Silicon.
Why It Matters: The Hybrid AI Value Proposition
SuperClaw exists to solve the agentic AI trilemma — cost, scale, and data security — and even in beta, the two-system setup is showcasing what hybrid compute unlocks:
- Dramatic cloud token savings: Agentic workloads are token-hungry: every tool call, every retry, every context replay burns cloud tokens. SuperClaw's intelligent routing keeps high-frequency tasks local and escalates only what genuinely needs frontier reasoning — demonstrating up to 70% reduction in average cloud compute token consumption on relevant enterprise workloads1. Local execution converts unpredictable cloud bills into fixed, amortized hardware you already own.
- Privacy protection, built in — not bolted on: Cloud-only agents offer no native protection for sensitive data. SuperClaw masks personally identifiable information on-device before anything reaches the cloud, detecting PII with 99% accuracy on industry-standard privacy benchmarks2. Your data is protected by architecture, not just by policy.
- Strong AI performance, without compromise: Across a broad set of hybrid routing benchmarks, SuperClaw matched or exceeded the task accuracy of cloud-only configurations. With an 80B-class model served on Arc Pro B70 GPUs at the edge, the beta delivers coding, cowork, deep research, and data-analysis experiences that feel every bit frontier-grade.
The Pinchbench results provide a great snapshot of SuperClaw’s potential, outputting answer quality on-par with similar cloud-based AI solutions while keeping 77% of the tokens on-device. Additionally, SuperClaw accomplished this with 100% accuracy in task delegation, ensuring each task was correctly delegated to the relevant sub-agents for all 116 tasks included in the benchmark3.
For more information on Intel SuperClaw, the video here provides a quick overview of its capabilities and how it can help reduce token compute costs, protect sensitive data, and scale seamlessly for enterprises looking to integrate agentic AI operations:
We’re very excited about this initial public beta release and have been very encouraged by positive feedback we’re hearing from customers in our collaborations so far. We can’t do this journey alone, and we’re grateful to all of our OEM customers and partners that are helping Intel SuperClaw realizing its full potential in this new era of Agentic AI.
And stay tuned for our upcoming Intel SuperClaw AI PC standalone beta release, which will be supported on Intel® Core™ Ultra Series 3-powered systems.
Additional information
- Solving the Agentic AI Trilemma: Cost, Scale, and Data Security (Intel SuperClaw Overview)
- Intel SuperClaw Github page: https://github.com/intel/intel-ai-builder/tree/main/superclaw
- Intel AI Builder Platform home page: https://aibuilder.intel.com/
- Intel Arc Pro home page: https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/workstations/b-series/overview.html
Disclaimers
Performance varies by use, configuration and other factors. Learn more at www.intel.com/PerformanceIndex.
AI features may require software purchase, subscription or enablement by a software or platform provider, or may have specific configuration or compatibility requirements. Data latency, cost, and privacy advantages refer to non-cloud-based AI apps. Learn more at intel.com/AIPC.
SuperClaw is built based on the OpenCode framework, with additional hybrid AI capabilities, privacy controls, local context management, model routing, governance, and platform optimization developed by Intel.
1Token consumption benchmark testing based on combination of table indexing and query tools/skills workloads. Cloud LLM based on GLM-5 model available on OpenRouter: https://openrouter.ai/z-ai/glm-5. Local LLM based on quantized Qwen 3.6-35B-A3B model - https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF/blob/main/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf – served with llama.cpp with “thinking mode” set to “off” on the Intel Core Ultra Series 3 processor. Testing was conducted on Intel Core Ultra X7 358H system with Intel Arc B390 built-in GPU and 64GB of memory running on Microsoft Windows 11 Pro. Results as of May 9, 2026.
2PII detection accuracy benchmark testing based on 20-category open-pii-masking-500K-ai4privacy dataset: https://huggingface.co/datasets/ai4privacy/open-pii-masking-500k-ai4privacy/viewer. Testing was conducted on an Intel Core Ultra X7 358H system with Intel Arc B390 built-in GPU and 64GB of memory running on Microsoft Windows 11 Pro. Results as of May 8, 2026.
The PII detection accuracy testing yielded an F1 score of 95%. An F1 score is a metric combining both precision and recall performance into one score – on a scale of 1-100 – with a higher score indicating better performance.
3 Results based on Pinchbench benchmark test consisting of 116-task OEM subset across nine (9) categories: analysis, coding, csv_analysis, log_analysis, meeting_analysis, memory, productivity, research, and writing.
Local LLM testing based on Qwen3-Coder-Next model, utilizing “mode llmrouter-local” for Arm A and “mode llmrouter (auto)” for Arm C: https://huggingface.co/Qwen/Qwen3-Coder-Next. Cloud-based LLM testing based on Cloud z-ai/glm-5 model, utilizing “mode llmrouter-cloud” for Arm B and “mode llmrouter (auto)” for Arm C: https://docs.z.ai/guides/llm/glm-5. L2 grading based on DeepSeek-V3.1 model: https://huggingface.co/deepseek-ai/DeepSeek-V3.1.
Testing was conducted on Intel Xeon Silver 4410Y (2x CPU) system with 512GB of memory and Intel Arc Pro B70 (4x GPU) with 32GB VRAM for each GPU, running on Linux Ubuntu 24.04.4 LTS operating system (kernel 6.14.0-1011-intel). Test results as of June 25, 2026.
Test metric definitions:
L1: Pinchbench Level 1 score. Delegation correctness: whether the parent agent routed each task to the correct sub-agent (local-file-agent, websearch-agent, email-agent, or build). Shown as tasks passed out of 116 and the percentage. Measures routing, not answer quality.
L2 mean: Pinchbench Level 2 score. Answer quality rated 0 to 1 by an LLM. We scored twice per task (4 times if the two runs differ by more than 0.2), then averaged across all 116 tasks. Higher is better.
L2 std: Standard deviation of the per-task L2 scores across the 116 tasks. Measures how consistent the quality is within one arm. Lower means more uniform.
Total tokens: All tokens used by the arm (prompt, completion, reasoning, cache-read) across the parent agent and every sub-agent session. For the router arm it equals Local plus Cloud.
Local tokens: Tokens processed by the on-device model. Here we use Qwen3-Coder-Next.
Cloud tokens: Tokens processed by the cloud model. Here we use z-ai/glm-5.
L2 / M cloud tokens: Quality per cost: L2 mean divided by (Cloud tokens / 1,000,000). Higher means each million cloud tokens buys more quality.