Pokee's 28B model runs 10m-token agents in-house. Published · Aug 9, 2026 Why Relve is watching this A small model that holds 10M tokens of context and runs entirely inside a customer's own boundary unlocks agentic deployments that regulated and data-sovereign teams couldn't do before, a real capability, even if the headline benchmarks are self-reported. Key Takeaways * Pokee AI released Pokee-Isaac 28B, a 28-billion-parameter agentic model with a 10-million-token context window, roughly 10x the typical 1M limit of mainstream models. * It's built to run inside the customer boundary: in a private cloud, on-premises, or on-device, so data never leaves the organization's control. * Pokee says it runs on a single GPU starting from a consumer RTX 4090, and scored 93.3 on the RULER long-context test at the full 10M-token length in its own testing. * API pricing is $0.15 per million input tokens and $1.00 per million output, with an OpenAI-compatible API. * Pokee was founded in 2024 by Zhu Zheqing, former head of Meta's applied reinforcement learning, and raised $12 million from Point72 Ventures, Qualcomm Ventures, and Samsung NEXT. What happened. Pokee AI released Pokee-Isaac 28B, a 28-billion-parameter agentic model with a 10-million-token context window, and positioned it around one idea: running long-context agents entirely inside a customer's own infrastructure. The model deploys in a virtual private cloud, on-premises, or on-device, so data never crosses an external boundary. That framing targets a real constraint. Long-context agentic capability has been almost entirely cloud-only because serving it cheaply anywhere else was hard, which locks out regulated industries, public-sector deployments, and any organization that legally can't send data outside its walls. A 28B model small enough to run on a single GPU changes what those teams can deploy. Pokee says Isaac runs on a single GPU starting from a consumer RTX 4090, and reports scoring 93.3 on the RULER long-context benchmark at the full 10-million-token length, where the company says other models failed beyond 2 million tokens. It exposes an OpenAI-compatible API priced at $0.15 per million input tokens and $1.00 per million output, and connects to more than 90 integrations via one-click OAuth. The model is currently text-only, with no image, audio, or video support. The company was founded in 2024 by Zhu Zheqing, former head of Meta's applied reinforcement learning, and raised $12 million in seed funding from Point72 Ventures, Qualcomm Ventures, and Samsung NEXT. It builds on the broader push to make agentic AI cheaper to run at scale. Why it matters. The genuine unlock here isn't the benchmark score, it's the deployment boundary. For engineering and ops teams in regulated industries, government, or any data-sovereign environment, a capable long-context agent that runs entirely in-house removes a hard blocker that cloud-only models can't clear, and it eliminates variable per-token pricing on data the organization already holds, a different value proposition from the cheapest-cloud-model race. The headline claims are almost all self-reported, and worth reading skeptically. The technical report names the architecture "non-decoder-only" but never explains what that means across 19 pages, some weights were fine-tuned from Qwen, and the flagship throughput figures come from a datacenter B200, not the consumer RTX 4090 in the marketing. On independent-style comparisons it places second on Terminal-Bench behind GPT-5.6 Luna, strong for its size, but not the across-the-board leader the launch framing implies. By keeping execution strictly within customer-governed boundaries, Isaac preserves complete data sovereignty and eliminates variable per-token pricing. Pokee-Isaac 28B Technical Report Bottom line. Watch whether independent evaluations confirm the 10M-token retention claims outside Pokee's own testing, and whether the architecture details get explained, since an unexplained "non-decoder-only" design is hard to build production dependency on. The in-boundary deployment model is the durable idea here, and it fits the broader shift in where AI inference economics are heading. For engineering and ops teams with data-sovereignty requirements, per Relve, an AI tools intelligence platform, Isaac is worth a controlled evaluation if you need long-context agents on data that legally can't leave your environment, just validate the long-context claims on your own workload before depending on the 10M-token figure. Neelam Khan. Lead Editor Neelam Khan is a Lead Editor at Relve, covering AI news, tools, product updates, search trends, and business use cases. She filters noise from useful signals for founders and teams, drawing on her previous work in AI SEO, content strategy, and tool research with Wellows and AllAboutAI.

Memories.ai, founded by former Meta engineers Dr Shawn Shen and Enmin "Ben" Zhou, has raised $16 million in seed funding to develop AI systems with persistent visual memory. Susa Ventures led the initial $8 million round in July, with Samsung Next, Seedcamp and others participating. An $8 million extension followed in March. The San Francisco startup has built a Large Visual Memory Model that enables AI to retain and search millions of hours of video footage. Early customers include physical security firms querying CCTV archives and marketing teams analysing brand mentions across social video. The company has indexed over 10 million hours of video and is working with Qualcomm and NVIDIA to bring the technology to edge devices, including wearables and robots, targeting applications beyond cloud-based analytics.

Drafted raises $16 million to make AI-driven home design accessible, backed by investors like Y Combinator and OneRepublic's Ryan Tedder.