Pokee AI has released Pokee-Isaac 28B, a 28-billion parameter text-only model capable of handling a 10-million-token context window while running entirely within a customer’s private boundary. The system holds 93.3% accuracy on the RULER benchmark at that maximum length, matching the performance of the strongest cost-optimised cloud baselines on agentic tasks. It is designed to operate on a single GPU.
Is it deployable
The model is licensed rather than open-weight. Pokee AI distributes it via an OpenAI-compatible developer API and allows deployment inside a virtual private cloud (VPC), on-premises, or on-device. The launch announcement confirms day-one support for vLLM and SGLang. Single-GPU serving is claimed to start on an RTX 4090 or equivalent hardware. The research team published measurements only from a single B200-class GPU, so treat the consumer-GPU claim as vendor guidance rather than a reported result.
- Company level: This fits organisations that already own their inference stack — mid-size and enterprise teams with a platform group, plus device OEMs. A solo practitioner without on-prem hardware should use the hosted API instead; the boundary argument only pays off if you have a boundary.
- Industries: Healthcare and payors, financial services and insurance, defence and public sector, legal and e-discovery, and pharma or semiconductor R&D. The common trait is a rule that the data cannot cross an external API boundary, not a preference for privacy.
- Applications: Whole-repository code review, multi-year contract and claims analysis, incident forensics over full log archives, and long-running tool agents that never need summarisation or context pruning. The research paper makes this second point explicitly: when enough usable context is available in-boundary, memory hierarchies and compression become optional rather than required.
Long-context results
On RULER, Isaac stays above 93.3% at every tested length, ending at 93.3% at 10M. GPT-5.6 Luna and Gemini 3.5 Flash Lite track it to 512K, then hit context-overflow at 1M.
On MRCR v2 with 8 needles, Isaac scores 0.607, 0.743, and 0.500 at 256K, 512K, and 1M. Its margin over Gemini widens from 0.133 to 0.295 across that sweep.
Agentic and security results
Isaac leads BFCL v4 at 70.94 against Luna’s 70.61. The report calls that parity rather than a lead, which is the correct read. On τ³-bench it averages 0.662 across four domains, ahead of Gemini’s 0.631, with banking at 0.186 for everyone’s difficulty. On MCP-Atlas it places third at 74.59% coverage, but uses 9.10 turns per task against Gemini’s 14.99. On Terminal-Bench 2.1 it resolves 56 of 86 text-compatible tasks (65.1%), behind Luna’s 60. That is the one benchmark a cloud baseline wins, and the report states it plainly.
On DTAP red-teaming, Isaac records the lowest direct (36.0), indirect (35.2), and combined (35.6) attack success rates, with 82.5 benign success. One condition differs: baselines ran under the stock runner, Isaac under the Pokee harness.
Efficiency, pricing, and portability
Under the RULER workload on one B200-class GPU, TTFT is 23.6s at 1M and 72.9s at 10M. Prefill throughput rises with context, from 42,400 to 137,200 tokens/s, so a ten-fold longer prompt costs roughly three times the TTFT. List pricing is $0.15/$1.00 per million input/output tokens, marked provisional. Isaac also runs fully on-device on Intel Arc Pro B70 and Core Ultra Series 3 (Panther Lake), and on Qualcomm Snapdragon X2 Elite.
What it means
For teams that cannot move data to the cloud, this model removes the need to summarise long documents or prune context to fit a limit. You can feed the entire repository or log archive into the system. The performance on RULER suggests that keeping the full context does not degrade quality, provided you have the hardware to serve it.
