Pokee-Isaac 28B: 10M-Token Context Window for On-Prem AI Deployment

Pokee AI launched Pokee-Isaac 28B on August 8, 2026, a 28-billion-parameter text-only model with a 10-million-token context window designed specifically to run inside a customer's own infrastructure. The model targets regulated industries where data cannot cross an external API boundary, and is licensed for deployment inside a VPC, on-premises, or on-device. It is not open-weight. Benchmarks show it holds 93.3% on RULER at 10M tokens, while competing baselines like GPT-5.6 Luna and Gemini 3.5 Flash Lite run out of context beyond 2M tokens.

What happened

Detail	Fact
Model name	Pokee-Isaac 28B
Parameter count	28 billion
Context window	10 million tokens
RULER score at 10M tokens	93.3%
API pricing (provisional)	$0.15 / $1.00 per million input / output tokens
Minimum GPU for serving	RTX 4090 or equivalent (vendor guidance)
Benchmark GPU used in research	Single B200-class GPU
Deployment modes	VPC, on-premises, on-device, hosted API
On-device hardware	Intel Arc Pro B70, Intel Core Ultra Series 3 (Panther Lake), Qualcomm Snapdragon X2 Elite
Inference framework support	vLLM, SGLang (Day-0)

Pokee AI positions Isaac around one specific problem: long-running AI agents accumulate tool outputs, observations, and reasoning steps faster than most context windows can hold them. Until now, the models capable of handling that scale of context have all been cloud endpoints, which rules out healthcare, financial services, defense, legal, and pharma or semiconductor R&D, where the data literally cannot leave the organization’s boundary.

Isaac is accessed through an OpenAI-compatible developer API, so existing code targeting GPT-style endpoints can route to it with minimal changes. For teams that already own inference hardware, Pokee licenses the model for self-hosted deployment.

How does Pokee-Isaac 28B perform on benchmarks?

The long-context story is the clearest part of the benchmark report. On RULER (a test of retrieval and coherence across very long inputs), Isaac stays above 93.3% at every tested length, including at 10M tokens. GPT-5.6 Luna and Gemini 3.5 Flash Lite both follow it to 512K but hit their context ceiling at 1M, returning 0.0 beyond that point.

On MRCR v2 with 8 needles (a multi-document retrieval test), Isaac scores 0.607 at 256K, 0.743 at 512K, and 0.500 at 1M. Its margin over Gemini grows from 0.133 to 0.295 across that range.

Agentic results are more mixed:

• BFCL v4 (function-calling): Isaac scores 70.94 vs. Luna’s 70.61. The Pokee team calls this parity, and that is the right description.
• t3-bench (multi-domain task completion): Isaac averages 0.662 across four domains, ahead of Gemini’s 0.631.
• MCP-Atlas: Isaac places third at 74.59% task coverage but uses 9.10 turns per task vs. Gemini’s 14.99 turns.
• Terminal-Bench 2.1: Isaac resolves 56 of 86 compatible tasks (65.1%), behind Luna’s 60. This is the one benchmark a cloud baseline wins, and the report says so plainly.

On the DTAP red-teaming evaluation (measuring resistance to adversarial prompt attacks), Isaac records the lowest direct (36.0), indirect (35.2), and combined (35.6) attack success rates, while maintaining an 82.5% benign task success rate. One caveat: the baselines ran under the stock test runner, Isaac under the Pokee harness, so the comparison is not perfectly controlled.

Serving performance

All published throughput numbers come from a single B200-class GPU. On the RULER workload, time-to-first-token (the delay before output begins) is 23.6 seconds at 1M tokens and 72.9 seconds at 10M. Prefill throughput (how fast the model processes the input) rises from 42,400 tokens per second at shorter contexts to 137,200 tokens per second at 10M, meaning a ten-fold longer prompt costs roughly three times the wait. Decode speed holds near 335 tokens per second.

Pokee also states the model runs on consumer-grade hardware starting from an RTX 4090. The research paper does not include measurements from consumer GPUs, so treat that as vendor guidance rather than a tested result.

Why it matters

The gap this fills is real. Regulated industries have been stuck choosing between capable cloud models they cannot legally use on sensitive data, or smaller open-weight models that run out of context window mid-task. A 10M-token window means a model can ingest an entire code repository, a multi-year contract archive, or a full incident log without summarization or context pruning.

The Pokee research paper makes this explicit: when sufficient context is available inside the boundary, memory hierarchies and compression pipelines become optional rather than required. That matters for agentic workflows where information lost to compression can cause incorrect decisions later in the task.

For teams exploring AI integration in regulated environments, this is a model worth watching, particularly because it targets the exact use cases where most SaaS AI tools are off the table by default.

Our take

The benchmark numbers are credible where they are specific, and Pokee deserves credit for being transparent about where a cloud model wins (Terminal-Bench 2.1) rather than burying it. The DTAP red-teaming caveat (different runners for Isaac vs. baselines) is worth noting before anyone uses that result for a compliance argument.

The bigger question is who this actually fits. The model is not open-weight, so you are taking on a vendor dependency even for on-prem deployment. The solo practitioner or small team without existing GPU infrastructure gets little advantage here. The value is real for organizations that already operate a private inference stack and need context lengths that cloud APIs simply cannot reach.

We have written before about how quickly AI inference costs compound at scale. At $0.15 input and $1.00 output per million tokens (provisional), Isaac’s hosted API sits in a reasonable range, but output token costs add up fast in multi-step agentic workflows. Run the numbers against your actual task profile before assuming on-prem licensing is cheaper than the API.

What to do about it

1. Audit whether your data actually cannot cross an API boundary by rule, not just by preference. If it cannot, Isaac is worth a pilot.
2. Check your current GPU inventory against the RTX 4090 minimum guidance before assuming you need new hardware.
3. Request the licensed weights or API access through Pokee’s developer API, which is OpenAI-compatible, so integration work is minimal.
4. Test your specific long-context task (repository review, contract analysis, log forensics) rather than relying on RULER scores. RULER measures coherence; your task may have different failure modes.
5. If you need help structuring an on-premises AI integration, map the data boundary requirements first, then choose the deployment mode.

The 10M-token context window is technically impressive. The more useful question is whether your specific task actually needs 10M tokens, or whether a well-structured smaller context would do the same job.

Source: Marktechpost