A hybrid design: pretrained models plus a custom network.

Nero uses pretrained language models for fluent conversation and code, and a separate custom network to maintain its internal state — memory, emotion, and continuity.

        +---------------------------------------+
 user ->|  mind.py - orchestration + router     |
        +-------------------+-------------------+
                            |  emotional state
                            v
        +---------------------------------------+
        |            HybridNero                  |
        |  Qwen2.5-Instruct    language model    |
        |  Qwen2.5-Coder       code model        |
        |  BiologicLLMV2 400M  internal network  |
        +---------------------------------------+
                            |
                            v   a single reply

The router

A router in mind.py reads each message and directs it to the right model — code requests to the code model, everything else to the language model. The current emotional state is passed to both as context.

The internal network

BiologicLLMV2 does not generate text. It maintains Nero's internal state: memory embeddings, contradiction detection, and weights that update on every interaction.

First-message coherence

Because language comes from a pretrained model, Nero responds coherently from the first message, with no training loop. The full system runs on a single free T4 GPU in about 8 GB of memory.

Code generation

Nero routes coding requests to the code model. When idle, it can also generate small programs on its own and save them to nero_creations/.

Autonomous execution is opt-in. Generating and saving code is always safe, since it is only text until explicitly run.

Sandboxed execution

Code is screened before running. Only a whitelist of pure modules is allowed; anything touching files, network, the OS, subprocesses, or introspection is rejected.
Approved code runs in an isolated subprocess with a timeout and capped output.
The sandbox was tested against known escape techniques, which were identified and closed.

Identity

Nero maintains a consistent first-person identity as a system running on hardware. This framing is applied throughout the codebase so responses stay in character and the system does not default to generic “as an AI” disclaimers.

Next: The Soul Voice