---
license: apache-2.0
base_model: google/gemma-4-12B-it
pipeline_tag: text-generation
tags:
- nvfp4
- toolkoder
- toolkode
- local-captain
- agentic-coding
- gemma4
---

# ToolKoder 12B NVFP4

ToolKoder 12B NVFP4 is the GPU-oriented local-captain runtime lane. It is
planned as the sibling of `toolkit-cli/toolkoder-12b-mlx`, exported from the
same private merged checkpoint after RC 0.1 blessing.

RC note: this card is staged documentation. The final NVFP4 runtime is not
published until the private merge, exact export, Toolkit R2 head-verified
manifest, runtime smoke, ToolKode telemetry, and adversarial blessing all pass.

## Intended Role

ToolKoder 12B is the local captain lane. It is designed to plan, route, review,
and accept coding-agent work while dispatching bounded work to local or frontier
workers.

| Field | Value |
| --- | --- |
| Public repo | `toolkit-cli/toolkoder-12b-nvfp4` |
| Sibling repo | `toolkit-cli/toolkoder-12b-mlx` |
| Model line | ToolKoder |
| Origin family | Gemma 4 |
| Runtime | NVFP4 |
| Primary role | Local/GPU coding captain |

## Release Policy

The public artifact may contain only quantized runtime files, sanitized model
metadata, tokenizer/config files, license text, and origin provenance.

No private SFT rows, private prompts, eval predictions, execution traces,
holdouts, LoRA adapters, full-precision roots, or BF16 roots belong in this
repository. BF16/full-precision merged checkpoints are private Toolkit R2
artifacts only.

## Finalization Path

1. Freeze the curated Gemma 4 12B SFT inventory.
2. Train and merge LoRA into a private higher-precision checkpoint.
3. Export MLX and NVFP4 variants from that same checkpoint.
4. Upload the exact exports to Toolkit R2 with head verification.
5. Run runtime smokes and local-captain ToolKode telemetry.
6. Run adversarial release review over the exact public staging folders.
7. Publish only after `release_ready` and `blessed` are true.

## Claims

No benchmark claims are made from this staged card. Public quality claims require
the final exported artifact and reproducible telemetry for that exact artifact.