Processing multi-million-token datasets no longer requires sending enterprise data across external cloud APIs. To allow execution strictly within customer boundaries, Pokee AI launched a 28B-parameter text-only foundation model called Pokee-Isaac 28B featuring a 10M-token context window.
According to benchmark results published by MarkTechPost, Pokee-Isaac 28B achieved a 93.3% score on the RULER benchmark at a 10M-token context length. The model is distributed under a commercial license for virtual private cloud (VPC), on-premises, or on-device deployments, alongside an OpenAI-compatible API.
Pokee AI asserts that processing 10 million tokens directly inside local infrastructure renders memory hierarchies and context compression optional for long-running autonomous agents, shifting long-horizon AI workloads away from third-party cloud platforms.
High-Throughput Local Execution and Multi-Domain Benchmarks
Pokee-Isaac 28B supports single-GPU serving with Day-0 integration for vLLM and SGLang. Performance testing on a single B200-class GPU showed prefill throughput rising with context length, reaching 42,400 tokens/s at 1M tokens and 137,200 tokens/s at 10M tokens. Decode throughput remained constant at approximately 335 tokens/s across all context sizes.
When processing prompts scaled to 10M tokens, the time-to-first-token (TTFT) reaches 72.9s, nearly tripling the 23.6s TTFT observed during 1M tokens testing. On functional benchmarks, Pokee-Isaac 28B scored 70.94 on BFCL v4, slightly edging out GPT-5.6 Luna at 70.61. On the τ³-bench four-domain average, Pokee-Isaac 28B scored 0.662—led by a 0.912 score in telecom—compared to Gemini 3.5 Flash Lite at 0.631, while all tested models scored 0.186 in the banking domain.
In Terminal-Bench 2.1, Pokee-Isaac 28B resolved 56 of 86 text-compatible tasks for a 65.1% success rate, trailing GPT-5.6 Luna’s 60 tasks but surpassing competing open-weight models. For needle-in-a-haystack retrieval on MRCR v2 with eight needles, Pokee-Isaac 28B scored 0.607 at 256K, 0.743 at 512K, and 0.500 at 1M tokens, expanding its lead over Gemini 3.5 Flash Lite from 0.133 at 256K to 0.295 at 1M tokens. On MCP-Atlas, Pokee-Isaac 28B achieved a 70.24 coverage score, averaging 9.10 turns per task compared to Gemini’s 14.99 turns, while reaching 74.59% coverage in broad evaluations.
| Tool | Key difference | Best for |
|---|---|---|
| Pokee-Isaac 28B | In-boundary deployment with 10M local context window | Strict data boundary compliance and on-premises agents |
| GPT-5.6 Luna | Higher task completion rate on terminal benchmarks (60/86) | Complex command-line reasoning via cloud API |
| Gemini 3.5 Flash Lite | Requires 14.99 average turns on MCP-Atlas tasks | Lightweight cloud agent workflows without boundary limits |
Enterprise Licensing Costs and Hardware Bottlenecks
During DTAP red-teaming safety evaluations using a custom “Pokee harness”—unlike baseline systems evaluated via a stock runner—Pokee-Isaac 28B registered a 35.6% combined attack success rate, consisting of a 36.0% Direct ASR and a 35.2% Indirect ASR. Among all evaluated systems, it achieved the top safety performance regarding attack resistance while securing an 82.5% benign task success rate (BSR). On-device deployment extends to hardware like Intel Arc Pro B70, Core Ultra Series 3 (Panther Lake), and Qualcomm Snapdragon X2 Elite, alongside vendor guidance citing compatibility down to consumer RTX 4090 GPUs.
Under provisional rates, Pokee-Isaac 28B is charged at $0.15 per million input tokens along with $1.00 per million output tokens. Conversely, enterprise buyers pay $0.40 input and $1.80 output for GPT-5.6 Luna, whereas rates for Gemini 3.5 Flash Lite run $0.30 input and $2.50 output per million tokens.
Executing 10M tokens inside enterprise boundaries shifts operational costs from cloud billing to local hardware. A 72.9s time-to-first-token places heavy demands on real-time agent responsiveness, and while the vendor points to RTX 4090 support, official benchmark performance was derived exclusively from enterprise-grade B200 GPUs.
📊 Key Numbers
- RULER benchmark at 10M tokens: 93.3% score (Pokee-Isaac 28B)
- BFCL v4 benchmark score: 70.94 (vs. GPT-5.6 Luna at 70.61)
- Terminal-Bench 2.1 resolved tasks: 56 of 86 tasks / 65.1% success rate (vs. GPT-5.6 Luna’s 60 of 86)
- τ³-bench 4-domain average score: 0.662 (with 0.912 in telecom and 0.186 across models in banking)
- B200 prefill throughput: 42,400 tokens/s at 1M tokens; 137,200 tokens/s at 10M tokens
- B200 decode throughput: ~335 tokens/s (stable across context lengths)
- Time-to-first-token (TTFT): 23.6s at 1M tokens; 72.9s at 10M tokens
- MCP-Atlas coverage and turn efficiency: 70.24 score at 9.10 turns per task (vs. Gemini’s 14.99 turns)
- DTAP red-teaming attack success rates: 36.0% Direct ASR, 35.2% Indirect ASR, 35.6% combined ASR
- Benign task success rate (BSR): 82.5%
- Provisional pricing: $0.15 per 1M input tokens / $1.00 per 1M output tokens
- MRCR v2 (8 needles) retrieval: 0.607 at 256K, 0.743 at 512K, and 0.500 at 1M tokens
🔍 Context
Benchmark evaluations and safety measurements were reported by MarkTechPost based on DTAP red-teaming and standard agentic benchmark suites. The release directly addresses the strict data sovereignty constraints faced by regulated industries that cannot pass entire internal databases through external APIs. This development accelerates the operational trend toward localizing massive-context LLM execution within customer firewalls. Unlike cloud rivals such as GPT-5.6 Luna, Pokee-Isaac 28B operates natively inside customer VPCs and on-premises environments while reducing input token costs. This release responds directly to hardware availability and enterprise demands for in-boundary long-context agent processing.
💡 AIUniverse Analysis
Our reading: The technical achievement lies in achieving high context recall across 10 million tokens while running inside local data boundaries on a single GPU. Escalating prefill throughput to 137,200 tokens/s at 10M tokens allows organizations to insert massive repositories directly into prompt context, removing the need to manage complex vector databases, chunking strategies, or context compression layers.
However, local execution introduces major hardware and latency trade-offs. Waiting 72.9 seconds for the first token creates severe bottlenecks for interactive agentic workflows that require rapid multi-turn decisions. Furthermore, Pokee-Isaac 28B is distributed under a proprietary commercial license rather than an open-weight release, limiting deep model inspectability and creating software vendor reliance. Additionally, advertising consumer GPU compatibility while publishing benchmark figures recorded only on high-end B200 hardware masks the actual memory bandwidth required for effective enterprise deployment.
For this architecture to become standard in 12 months, Pokee AI must prove that consumer-grade GPUs like the RTX 4090 can process multi-million-token contexts without prohibitive latency degradation.
⚖️ AIUniverse Verdict
👀 Watch this space. Pokee-Isaac 28B demonstrates an exceptional 93.3% score on the RULER benchmark at 10M tokens, but its 72.9-second time-to-first-token latency and unverified performance on non-B200 hardware demand thorough real-world validation.
🎯 What This Means For You
Founders & Startups: Founders building for regulated sectors can deploy long-context agentic workflows without routing sensitive customer data through cloud API providers.
Developers: Developers can build repository-scale code analysis and persistent tool agents without needing to implement memory compression or context pruning layers.
Enterprise & Mid-Market: Enterprise platform teams with on-premises hardware can run massive-context agentic models in-house while maintaining strict regulatory compliance.
General Users: End users in healthcare, legal, and finance gain access to deep historical document analysis without having their sensitive data leave localized boundaries.
⚡ TL;DR
- What happened: Pokee AI released Pokee-Isaac 28B, a 28B-parameter text model offering a 10M-token context window for in-boundary enterprise deployment.
- Why it matters: It allows long-running agents to execute locally inside virtual private clouds without relying on external cloud APIs or context compression.
- What to do: Audit on-premises GPU memory bandwidth and test time-to-first-token latency tolerances before replacing cloud model endpoints.
📖 Key Terms
- RULER benchmark
- A synthetic evaluation framework designed to test long-context retrieval and needle-in-a-haystack accuracy across multi-million token context windows.
- MRCR v2
- A multi-step reasoning and context retrieval benchmark used to measure information extraction precision across expanding context lengths.
- vLLM
- An open-source library designed for fast and memory-efficient LLM inference and serving.
- SGLang
- A high-performance execution engine designed to accelerate structured generation and complex agent workflows on LLMs.
Analysis based on reporting by MarkTechPost. Original article here.