Built on Google's Gemma 4 lineage, this variant is delivered as a multimodal mixture-of-experts model in a 26B / A4B configuration that has been further tuned under the "Heretic" recipe to lower refusal rates while keeping coding, reasoning, and function-calling intact. It is published as an open-weight model for self-hosted deployments and runs inside a Trusted Execution Environment with provider attestation, giving operators a verifiable hardware-rooted boundary around inference. The NanoGPT listing emphasizes efficient chat behavior alongside the tooling and structured-output affordances that are typical of the Gemma family, making it a drop-in option for developers who want Gemma-class multimodal understanding without the standard content guardrails.

In practice the model is well suited to assistants that need image-aware comprehension, code generation, and tool use within a 65,536-token context and output budget, and the TEE deployment model is useful for workflows that require proof that prompts and weights are handled in an attested environment. The reduced-refusal tuning makes it a better fit for research, red-teaming, and product prototyping where conventional safety filtering would block legitimate experimentation, while the preserved reasoning and coding strengths keep it useful for general assistant workloads. On NanoGPT it is offered only through auto routing, with no published benchmark suite, so users evaluating it should rely on their own task-specific tests rather than leaderboard scores.