# Pokee AI ships closed-source (for now) 28B model designed for 10-million-token workloads > Source: > Published: 2026-08-04 18:56:59+00:00 [Zheqing (Bill) Zhu (@ZheqingZhu)](https://x.com/ZheqingZhu), founder and CEO of [Pokee AI](https://pokee.ai/), launched Pokee-Isaac 28B on Tuesday, August 4th, pitching a proprietary agent model that can process 10 million tokens of context on a single Nvidia RTX 4090-class GPU. [In its announcement on X](https://x.com/Pokee_AI/status/2084682445648216383), Pokee said the 28-billion-parameter model scored 93.3% on RULER at the full 10-million-token length and reached prefill speeds of as much as 137,000 tokens per second on one GPU. Pokee described Isaac as a "frontier-class agentic model" built on a proprietary "non-decoder-only" architecture. Those performance figures come from Pokee's internal testing. The company has not released the current model's weights, and the launch material discloses only the architecture label rather than enough technical detail to assess how it reduces the memory and compute demands associated with very long prompts. Pokee said in replies to the announcement that it intends to open source a version in the future, without setting a date. The claimed context length is the center of the launch. Ten million tokens could allow an agent to inspect large document collections, code repositories, tool catalogs and conversation histories in one session without relying as heavily on retrieval systems that split data into smaller chunks. Pokee's website says Isaac is designed to plan, execute and review long-running tasks while operating inside a customer's infrastructure. RULER, the benchmark behind Pokee's headline score, is a synthetic evaluation that tests retrieval, multi-hop tracing and aggregation across long inputs. Its [original research paper](https://arxiv.org/abs/2404.06654) was designed to distinguish an advertised context window from the amount of context a model can use effectively. A strong RULER result supports Pokee's long-context claim, but it does not establish general reasoning quality or reliability in multi-step enterprise workflows. Pokee's own launch graphic also shows Isaac trailing GPT-5.6-Luna on Terminal-Bench 2.1 and MCP-Atlas while narrowly leading it on the company's reported BFCL v4 and tau3 averages. All comparisons in the graphic are marked as internal or sourced from external vendors, leaving the claimed 10-million-token performance and single-GPU throughput without independent replication. Pricing is inconsistent across Pokee's launch materials. The benchmark graphic advertises $0.15 per million input tokens and $1 per million output tokens. Pokee's [API documentation](https://docs.pokee.ai/docs/models), accessed on August 4th, lists the standard Isaac tier at $0.30 per million input tokens and $5 per million output tokens. A high-reasoning tier is listed at $3 and $15, respectively. ### Zhu's enterprise deployment bet Isaac extends the thesis Zhu brought to Pokee after leading applied reinforcement learning at Meta AI. According to [Zhu's public biography](https://www.zheqingbillzhu.com/home), he led Meta's Pearl production reinforcement-learning project and completed a Stanford PhD in reinforcement learning while working full time. He founded Pokee in 2024 around agents that plan and use tools across applications rather than stopping at generated text. That background shaped Pokee's initial product. The Seattle-area company applies reinforcement learning to tool selection and task sequencing across services including Google Workspace, Slack, GitHub and Notion. In July 2025, [GeekWire reported](https://www.geekwire.com/2025/point72-ventures-leads-12m-seed-round-for-a-new-ai-agent-startup-founded-by-ex-meta-manager/) that Pokee had raised a $12 million seed round led by [Point72 Ventures](https://p72.vc/), with participation from Qualcomm Ventures, Samsung NEXT and other investors. Pokee was pre-revenue at the time and operating a public beta. The single-GPU claim gives Zhu a sharper enterprise sales argument. Pokee offers hosted access alongside [private deployments](https://pokee.ai/deployment) in a customer's cloud account, and its agent infrastructure can run on-premise or in an air-gapped environment. A model that genuinely handles 10 million tokens on commodity high-end hardware would reduce dependence on large inference clusters and give companies tighter control over sensitive documents and workflow data. Pokee is offering Isaac through its API and custom deployments rather than releasing weights developers can inspect and run independently. That makes the launch an enterprise infrastructure play first: Pokee controls the model, inference service and agent runtime, while customers can keep the workload inside their network boundary. Isaac's commercial case now depends on whether outside developers can reproduce the headline performance on the hardware Pokee names. Until that happens, the 10-million-token RULER score, 137,000-token-per-second prefill rate and RTX 4090 deployment remain vendor-reported results.