**Pokee-Isaac 28B: A Long-Horizon, On-Premises Foundation Model**
Long-horizon agents accumulate context faster than they resolve tasks. Every tool output, observation, and intermediate reasoning step remains in the window, making two capabilities critical: the ability to hold vast context and stay coherent across it. Until recently, these features were almost exclusively available via cloud endpoints, disqualifying regulated industries, public-sector institutions, and on-device applications from such architectures.
Pokee AI addresses this gap with **Pokee-Isaac 28B**, a 28-billion-parameter text-only foundation model designed to operate with a **10M-token context window**. The model targets deployment inside a customer’s security boundary—on-premises, in a VPC, or even on-device—without transmitting data externally. Independent benchmarks claim **93.3% on RULER at 10M tokens**, matching the strongest cost-optimized cloud models on agentic benchmarks while fitting on a single GPU.
—
### Is It Deployable?
Pokee-Isaac 28B is **deployable but licensed, not open-weight**. The model is delivered through an OpenAI-compatible API and can run in VPC, on-premises, or on-device environments. Day-0 support is available for **vLLM and SGLang**, with single-GPU serving starting from an RTX 4090 or equivalent hardware.
– **Enterprise & device OEMs** with existing inference stacks are the primary targets.
– **High-regulation industries**—including healthcare, finance, defense, legal, and semiconductor R&D—are the main beneficiaries.
– Applications include **full-repository code review, multi-year contract analysis, log-forensics, and always-on tool agents** that never require summarization or context pruning.
> Note: A solo practitioner without on-prem hardware is better served by the hosted API, since the value proposition only matters when you control the boundary.
—
### Long-Context Performance
Isaac 28B maintains near-constant accuracy even as context scales to millions of tokens.
– **RULER scores remain above 93.3%** across all tested lengths, reaching 10M tokens.
– **MRCR v2 (8-needle)** tests show strong recall under interference, with scores of **0.607 (256K), 0.743 (512K), and 0.500 (1M)**.
– Unlike GPT-5.6 Luna and Gemini 3.5 Flash Lite—which collapse at 1M tokens due to context overflow—Isaac continues delivering usable results up to 10M tokens.
—
### Agentic and Security Results
Isaac matches or exceeds cloud models in key agentic benchmarks while demonstrating superior security posture.
| Benchmark | Score | Insight |
|———————–|——————-|——————————————-|
| BFCL v4 | 70.94 | Tied with top cloud baseline |
| τ³-bench (4-domain) | 0.662 avg | Leads Gemini 3.5 Flash Lite |
| Terminal-Bench 2.1 | 65.1% | Slightly behind Luna (69.8%), but strong |
| MCP-Atlas | 74.59% coverage | Third place; fewer turns than competitors |
| DTAP Red Teaming | 35.6% combined ASR| **Lowest attack surface**; 82.5% benign success |
In red-team testing, Isaac records the **lowest direct and indirect attack success rates**, maintaining high benign task completion without relying on refusal-only defenses.
—
### Efficiency, Pricing, and Portability
Isaac is engineered for efficient, boundary-compliant serving.
– **Latency**: 23.6s time-to-first-token at 1M context; **prefill throughput rises with context**, reaching **137,200 tokens/s** at 10M tokens.
– **Decode speed** remains stable at ~335 tokens/s.
– **Pricing**: Listed at **$0.15 per M input tokens** and **$1.00 per M output tokens** (provisional).
– **Hardware**: Runs on a single B200-class GPU and also supports **Intel Arc Pro, Core Ultra Series 3, and Qualcomm Snapdragon X2 Elite** for on-device deployment.
—
### Cost Comparison (Estimate)
| Model | Input Price ($/M) | Output Price ($/M) |
|—————————–|——————-|——————–|
| Pokee-Isaac 28B | $0.15 | $1.00 |
| GPT-5.6 Luna | $0.40 | $1.80 |
| Gemini 3.5 Flash Lite | $0.30 | $2.50 |
*Note: Cloud-based long-context rates apply only above certain token thresholds.*
—
### Key Takeaways
– **Stateful long-context performance**: 93.3% on RULER at 10M tokens; no baseline survives beyond 2M.
– **High throughput**: 137k prefill tokens/s at maximum context on one GPU.
– **Benchmark leadership**: Leads BFCL v4 and τ³-bench; competitive on Terminal-Bench and MCP-Atlas.
– **Security-first**: Lowest red-team attack success with high task success.
– **Deployment flexibility**: Licensed for VPC, on-prem, and on-device use; weights not公开.
—
### FAQ
**Q: Is Pokee-Isaac 28B open source?**
A: No. The model is licensed, not open-weight. Deployment requires a Pokee AI agreement.
**Q: Can it run on consumer hardware?**
A: Yes. Single-GPU serving starts from an RTX 4090 or equivalent.
**Q: What makes it suitable for regulated industries?**
A: Data never leaves the customer boundary—on-prem, VPC, or device—meeting strict compliance requirements.
**Q: How does it compare on long-context benchmarks?**
A: It outperforms all public cloud models at 1M+ tokens, avoiding context overflow issues.
**Q: What about pricing for on-prem deployment?**
A: Pricing is hardware-based and not token-metered, unlike cloud endpoints.
—
### Conclusion
Pokee-Isaac 28B represents a practical shift for organizations that require long-horizon reasoning without sacrificing data sovereignty. By combining a 10M-token context window with strong agentic performance and boundary-first deployment, it removes the traditional trade-off between capability and compliance. For enterprises operating in regulated environments—or those simply unwilling to move sensitive data to the cloud—Isaac offers a compelling, production-ready alternative to conventional cloud-native models.
