Company news and technical perspectives from INT21.https://int21.ai/https://int21.ai/insights/pushing-gpt-5-6-sol-to-100-on-arc-agi-3-public/https://int21.ai/insights/pushing-gpt-5-6-sol-to-100-on-arc-agi-3-public/SwarmOS pushed GPT-5.6-Sol from a 13.3% baseline to 100% RHAE on ARC-AGI-3 Public, showing how orchestration can multiply long-horizon agent capability.Thu, 27 Aug 2026 00:00:00 GMThttps://int21.ai/insights/addressing-the-inference-bottleneck/https://int21.ai/insights/addressing-the-inference-bottleneck/INT21's agent swarms built live inference engines for video, music, and speech generation—without hand-built frameworks or human expert tuning.Tue, 18 Aug 2026 00:00:00 GMThttps://int21.ai/insights/introducing-int21-and-ptx-kernel-factory/https://int21.ai/insights/introducing-int21-and-ptx-kernel-factory/We are building self-improving AI systems for the software beneath modern AI. Our first product generates and optimizes low-level GPU software, then proves its work with tests and benchmarks.Tue, 16 Jun 2026 00:00:00 GMThttps://int21.ai/insights/drama-3-shipped-one-hour-later-we-ran-it-63-percent-faster-than-day-zero-support/https://int21.ai/insights/drama-3-shipped-one-hour-later-we-ran-it-63-percent-faster-than-day-zero-support/INT21 generated a Rust inference runtime for Fish Audio's new speech model, Drama 3, in just one hour. Despite receiving no early access, we achieved 59–63% higher throughput and 38–60% lower time to first audio than SGLang-Omni (which supported Drama 3 on day one), in our single-stream benchmarks.Wed, 14 Oct 2026 00:00:00 GMThttps://int21.ai/insights/ai-generated-inference-engines/https://int21.ai/insights/ai-generated-inference-engines/Two engineers. Twenty engines. Two weeks. What our comparisons with SGLang and vLLM reveal about the promise—and the unfinished work—of AI-generated infrastructure.Mon, 28 Sep 2026 00:00:00 GMThttps://int21.ai/insights/swarmos-gpt-6-astra-qwen38-trainer/https://int21.ai/insights/swarmos-gpt-6-astra-qwen38-trainer/SwarmOS and GPT-6 Astra generated a Rust/CUDA Qwen3.8-27B trainer with 11.5× the throughput of eager PyTorch FSDP2 and 21.5% higher throughput than the tuned baseline.Mon, 07 Sep 2026 00:00:00 GMThttps://int21.ai/insights/what-is-self-improving-compute-infrastructure/https://int21.ai/insights/what-is-self-improving-compute-infrastructure/Self-Improving Infrastructure uses AI agent swarms to continuously improve the software AI runs on, making it faster, cheaper, and more reliable.Mon, 13 Jul 2026 00:00:00 GMThttps://int21.ai/insights/muse-spark-1-1-enterprise-agent-architecture/https://int21.ai/insights/muse-spark-1-1-enterprise-agent-architecture/We ran Meta's Muse Spark 1.1 inside SwarmOS. The result shows why the durable enterprise advantage is the orchestration layer, not any single model.Thu, 09 Jul 2026 00:00:00 GMThttps://int21.ai/insights/stop-waiting-for-a-bigger-context-window/https://int21.ai/insights/stop-waiting-for-a-bigger-context-window/Bigger context windows don't solve long context; multi-agent orchestration does. Here's how we tackle workloads involving tens of millions of tokens.Tue, 07 Jul 2026 00:00:00 GMT