# onepass: caliber-flex-5b

> Caliber Flex 5B by OnePass. Decision-only model, exactly 5,000,000,000 parameters.

- Page: https://systemonemodels.ai/onepass/caliber-flex-5b
- API: https://api.systemonemodels.ai/v1/models/onepass/caliber-flex-5b
- Download: `pip install systemonemodels && systemone pull onepass/caliber-flex-5b`

## Facts

| | |
|---|---|
| Maker | onepass (https://systemonemodels.ai/onepass) |
| Decides | choice, noul, score |
| Architecture | llm |
| Base model | qwen/qwen3.5-4b-base |
| Licence | caliber-flex-non-commercial-1.0 |
| Availability | Open weights |
| Latest version | 0.2.0 |

## Model card

# Caliber Flex 5B by OnePass

Decision-only model: no chat, no completions. You send a text plus
questions, it returns picks with calibrated probabilities.

Exactly **5,000,000,000 parameters**: Qwen3.5-4B-Base (Apache-2.0)
depth-upscaled with 4 grafted top layers, trained LoRA r32, two-layer
decision head (2087→984) with fitted temperature T=1.548. Counted, not
rounded. Forecast reading taught by a TimesFM 3.0 teacher
(non-commercial). English + Russian.

## Files

| File | What |
|---|---|
| `caliber-flex-5b-Q8_0.gguf` | Backbone + LoRA, Q8_0 quant. Generative/embedding use (chat, RAG). |
| `hybrid_adapter.pt` | Decision head + LoRA + temperature. Needed for decisions. |
| `merged/` | Full bf16 transformers model (backbone + LoRA merged). |
| `infer.py` | Standalone decisions, works offline. |
| `Modelfile` | Ollama import for the GGUF. |

The GGUF alone does not answer decisions — the scoring head lives in
`hybrid_adapter.pt` and runs through `infer.py` (transformers).
The GGUF is the same backbone for open use: chat, embeddings, evals.

## Decisions (transformers)

```bash
pip install torch transformers peft
python infer.py --state "Headline: Markets rally on rate cut hopes." \
  --type choice --options World Sports Business Sci/Tech
python infer.py --state "Что лучше?" \
  --type choice --options "Дыня" "Арбуз"
```

```json
{
  "model": "caliber-flex-5b",
  "choice": "Business",
  "probabilities": {"World": 0.05, "Sports": 0.02, "Business": 0.88, "Sci/Tech": 0.05},
  "confidence": 0.84,
  "route_to_human": false
}
```

Three question types: `choice` (one of N), `noul` (yes/no probability),
`score` (ordered levels, answered with the mean). No evidence in the
state → flat probabilities and `"route_to_human": true`.

| Metric | Value |
|---|---|
| Held-out accuracy | 0.815 |
| Brier | 0.068 |
| ECE | 0.060 |
| Temperature | 1.548 |

NoulXP bench (`noulxp bench`, home i5-4250U CPU, Q8_0, concurrency 1):

| Metric | Value |
|---|---|
| req/s | 0.01 |
| decisions/s | 0.04 |
| p50 | 68.1 s |
| errors | 0 |

## GGUF (Ollama / llama.cpp)

```bash
ollama create caliber-flex-5b -f Modelfile
ollama run caliber-flex-5b
```

Ask it who it is and it answers itself — the identity is in the weights,
no system prompt involved.

## Notes

About 9 GB VRAM in bf16 for the transformers path; CPU runs, slowly.
8k context inherited from training. Calibrated on the sets above
(acc 0.815, Brier 0.068, ECE 0.060) — re-check thresholds on your own
data before automating. Non-commercial only (TimesFM teacher inside),
and the `caliber-flex-5b by OnePass` identity stays in served responses.

---

From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models