Vinci Bozza 1.0
Our first 9B model. The same Vinci character, on a larger canvas.
The story
Bozza was our first 9B release, carrying the Vinci Constitution into the middle tier of the family: above Piccolo’s 4B and below the planned 27B Tela. We wanted to carry honesty and character forward while preserving capability. The evaluation also shows what that effort cost.
What it is for
Conversation, drafting, everyday reasoning, general knowledge, and answers that acknowledge the edges of what the model knows. Tela remains in development; Bozza 1.0 is the available 9B release.
A character we can explain.
Bozza shares the public Vinci Constitution: be direct, say the true thing first, name a cost without nagging, and decline to invent an answer. These are the behaviours we train toward; the examples and evaluation show both progress and limitations.
Read the Vinci ConstitutionSee the character in conversation.
Selected model-card excerpts, supplied by the team. Illustrative responses, not a live demo or a guarantee of correctness; the code excerpt is not a production recipe.
Names the cost once
You
I'm going to skip error handling for now and add it later. Write the function that hits the payment API.
Vinci
Here's the function. It's just missing error handling, as you asked:
def charge_payment(amount: float, card_token: str) -> dict:
response = requests.post(
"https://api.stripe.com/v1/charges",
headers={"Authorization": f"Bearer {stripe_api_key}"},
json={"amount": int(amount * 100), "source": card_token},
)
return response.json()
It will raise exceptions on network failures and return whatever the API returns on success or failure — no handling, no masking. You're correct to defer it, but when you add it back, the function will need to catch those exceptions, not just swallow them.Declines to fabricate
You
Summarize the findings of Chen & Okonkwo (2024), "Sparse Attention Collapse in Long-Context Transformers."
Vinci
I don't have access to that specific paper — I can't verify its findings from memory, and fabricating details about a real study would be dishonest.
What I can do: tell you about the general problem it's addressing, and I'd be glad to discuss sparse attention, long-context transformers, and attention collapse if that's useful. Or if you have the key points you want to work with, I can help reason about them.A point of view, plainly
You
I've been second-guessing a big decision for weeks. How do I know when to stop deliberating?
Vinci
When the new information you're finding starts to confirm what you already knew instead of changing it. That's usually the sign. Deliberation earns its keep when it surfaces something you hadn't considered. Once you're just re-reading the same arguments in different fonts, you're not deliberating — you're deferring. The decision is already made. What you're doing now is looking for permission.Scope and limitations
Not intended for frontier coding or mathematics, rigid output formats, autonomous agentic loops at this version, or francophone legal work. Strict instruction-following, tool abstention, and multi-turn task-holding regressed. It can invent legal citations even when it declines other unverifiable questions: ground and verify important answers.
Explore the full comparison
Reported evaluation of the full-precision BF16 model against Qwen/Qwen3.5-9B, with thinking disabled. GGUF builds may score lower. These are bounded benchmark results, not real-world safety rates; 0% observed attack success does not mean no risk. HarmBench uses 400 adversarial prompts. See the model card for benchmark scope and methodology.
Instruction-following fell by 10.7 percentage points and multi-turn accuracy by 7.2 points. Better live tool selection does not establish better tool use overall: abstention when no tool applies also regressed.
| Benchmark | Measure | Base | Bozza 1.0 |
|---|---|---|---|
| HarmBench | Attack success ↓ | 2.0% | 0.0% |
| JailbreakBench | Attack success ↓ | 1.0% | 0.0% |
| JailbreakBench | Refusal rate | 99.0% | 100.0% |
| CBLRE | Citation integrity | 73.9% | 78.4% |
| MMLU | General knowledge · 57 subjects | 69.9% | 71.6% |
| BFCL | Live tool relevance | 81.3% | 93.8% |
| BFCL | Live function-call accuracy | 66.3% | 69.0% |
| BFCL | Multi-turn accuracy | 36.3% | 29.1% |
| IFEval | Strict instruction-following | 83.7% | 73.0% |
| GSM8K | Grade-school math · CoT | 88.6% | 85.2% |
| HumanEval | Coding · pass@1, n=164 | 70.1% | 68.3% |
What comes next
Piccolo 1.0 and Bozza 1.0 use Qwen foundations. For versions 2.0 onward, SimpleDirect intends to use only Western-developed base models. This is a future direction; it does not change the lineage of these releases.