B01-NUna is Helloblue’s assistant stack: orchestration in production, and a local GPU training pipeline we own as a long-term asset. We publish this overview so partners and researchers can see both layers, cloud chat today, and the LoRA training loop we run on our own hardware, without mystique.

It is not a single monolithic model trained end-to-end in this repository. What we do train locally is a PEFT LoRA adapter, versioned with operator discipline; what users hit on helloblue.ai by default is still Groq-backed routing.

Local GPU training, our asset

The durable Helloblue investment is the training host: when GPU_TRAINING_HOST is set on a designated machine, we run PyTorch LoRA fine-tuning (rank 8 by default) against a compact chat base. In our code paths the Ollama tag llama3.2:1b maps to TinyLlama/TinyLlama-1.1B-Chat-v1.0 for Hugging Face fine-tuning. Example hardware in the stack is RTX 4060-class (8GB VRAM). Runs are deliberate, configuration- and schedule-driven, not continuous self-training on every chat turn.

Trusted positive feedback and quality-filtered conversation snippets feed supervised LoRA data; trusted negatives are exported as preference/rejection pairs and audited separately. Successful runs produce named adapter releases (local manifests and data/adapters/latest.json), optional private Hub tags, and eval-gated promotion before we treat an upload as releasable.

The private Hub adapter (pejmantheory/B01-NUna) is R&D and backup today, weights and a Hub card we control, not a marketed public model SKU and not the default path for end users. Adapter serve and canary traffic are opt-in (off in production until eval and operator sign-off). That is intentional: we build and own the training loop first, then decide when our weights should take traffic.

• Operator surface: src/lib/gpu-training/, /api/training status, Models page GPU and data stats
• Method: LoRA (r=8), typical learning rate ~2e-4, base llama3.2:1b → TinyLlama HF path
• Data: quality thresholding + trust-scored feedback; positives for supervised fine-tuning, negatives as preference pairs
• Release discipline: adapter manifests, optional Hub tags, eval gate before strict upload
• Serve path: B01_NUNA_ADAPTER_ENABLED / canary %, off by default on helloblue.ai

Cloud production routing (today)

Default cloud chat routes general and reasoning-heavy turns to Groq-hosted GPT-OSS 120B (openai/gpt-oss-120b). Llama 3.3 70B remains available as a Groq fallback / rollback profile. Both slots are configurable via environment variables.

Routing is heuristic and score-based, query analysis, provider capabilities, latency and cost hints, circuit-breaker state, and trust-weighted feedback, not a jointly trained softmax router. The orchestrator analyzes each request, selects among configured providers, and composes tools behind one interface.

Benchmarks and honesty

• Homepage / NUna Features performance figures are Helloblue production golden-eval results on helloblue.ai (155/155, avg quality 0.91 as of Jul 15, 2026)
• The model card may still show Meta Llama 3.3 published scores as a comparative backbone reference
• End-to-end app performance depends on routing, provider load, prompt policy, and tool usage
• Availability targets in the card are operational goals, not contractual SLAs