Pokee-Isaac 28B

At this stage, Pokee-Isaac 28B is best treated as a notable but still lightly verified long-context agent model, not a broadly validated mature foundation model. The available evidence is mostly official X posts, reposts, a technical report link, and ecosystem support claims. That is strong proof of attention, but weaker proof that it is already reliably good for your own workload. It is not a general consumer chatbot, and not simply a standard Qwen fine-tune; a better analogy is a deployable research-style model built around ultra-long context and agent behavior.

Based on the cited material, its headline claims are a 28B scale, “real” 10M-token context, single-GPU deployment, and a proprietary non-decoder-only architecture. Social posts also mention RULER results at 10M length, high single-B200 prefill speed, and operation starting from an RTX 4090. That suggests practical use in reading huge corpora, long-chain retrieval, tool use, and agent task orchestration for developers who want to stuff far more context into one run. Still, most of this is stated by the company or repeated by others, so these are capability claims rather than strong independent validation.