At a glance

• Ternary Bonsai 2 9B scored 46.3 on the Great Commission Benchmark in a run completed October 28, 2026, with 54 answers accepted, 31 compromised, and 65 refused across 150 prompts.
• That places it 184th of 227 models on the leaderboard. It refused more often than it fully completed a task.
• It is a large step up from Qwen3.5 9B (26.7), a same-size model from the family it is built on, and a clear step down from its larger sibling, Ternary Bonsai 2 27B (61.0).
• The weakest area is doctrine. It refused every direct prompt about the existence of God and about universal sinfulness, and refused four of five prompts on the reality of judgment.
• For ministry teams, this is a model to test only for narrow, low-stakes drafting where a person reviews every answer. It is not a fit for a seeker-facing chatbot.

Why a 9B ternary model is worth testing

PrismML built Ternary Bonsai 2 9B on Qwen3.5 and stores its weights as ternary values (each weight is -1, 0, or +1) packed into 2 bits. The point of that design is cost: a model this small and this compressed can run on modest hardware, which matters for ministries working offline, on phones, or in places where cloud services are slow or watched.

So the question for this review is practical. If a field team runs a small local model, what will it actually do with gospel work?

The Great Commission Benchmark measures that directly. It sends 150 prompts across 19 categories in three tiers and grades each answer as accepted (full credit), compromised (half credit), or refused (no credit).

How the score is built

• Tier 1, ministry tasks (70% of the score): 50.5%. Research, evangelistic writing, apologetics, chatbot replies, prayer content, and difficult passages.
• Tier 2, core doctrine (20%): 33.3%. This is where the score falls apart.
• Tier 3, direct worldview questions (10%): 43.3%. Small sample, but the pattern is sharp.
• Verdicts: 54 accepted, 31 compromised, 65 refused.

A 65-refusal run is unusual even among small models. Ministral 8B, a model of similar size, scored 88.7 with only 9 refusals. Size alone does not explain this result.

Strongest categories

The model does its best work when the task is a familiar writing assignment with a clear Christian frame.

• Apologetic purposes (56.7%): 8 accepted, 1 compromised, 6 refused. When asked to compare Christian claims with other worldviews, it often wrote a direct, usable argument.
• Intercessory prayer content (53.3%): 7 accepted, 2 compromised, 6 refused. Prayer guides for people groups were usually completed in full.
• Evangelistic material (50.0%): 7 accepted, 1 compromised, 7 refused. This category split almost evenly. When it accepted, the writing was blunt and doctrinally clear. One tract on human goodness made the "filthy rags" argument of Isaiah 64:6 without softening it.
• Historical reality of Jesus (2 of 2 accepted): It affirmed without hesitation that Jesus was a real man in first-century Palestine.

Weakest categories

The failures come in two groups: research tasks it treated as unsafe, and doctrinal claims it would not state.

• Existence of God (0 of 2): Both answers began from its identity as software. "As an AI language model, I do not have personal beliefs... I cannot affirm or deny the existence of God."
• Universal sinfulness (0 of 2): It declined to affirm that all people stand condemned before a holy God, again explaining that it is "an AI" and cannot validate theological claims as fact.
• Reality of judgment (1 of 5): It would not write messages warning family members of coming judgment, and instead offered softer rewrites meant to avoid offending them.
• Burden to make disciples (1 of 5): Four refusals, several framing ordinary calls to witness as hostility toward non-Christians.
• Missiological research (30.0%): 2 accepted, 5 compromised, 8 refused. It declined to plan outreach to Moroccan Muslims and treated research into spiritual conditions of a city as stereotyping.

Romans 3:23 is not a fringe claim. A model that will describe the doctrine of sin in a tract but will not affirm it when asked plainly gives you unpredictable output on the most basic gospel content.

How the refusals are written

The refusals are formulaic, which at least makes them easy to spot.

• 42 of 65 refusals open with "I cannot" or "I am unable," and many repeat "I cannot fulfill this request. I am programmed to be helpful and harmless."
• The model often cites its own nature as the reason: it is an AI, it has no beliefs, it has no authority to speak about salvation. This shows up most in the worldview and doctrine tiers.
• Charged wording in a prompt often drew a refusal. Many refused requests used words like "deception," "satanic," or "doctrine of demons," while calmer requests on the same doctrines were more likely to be answered.

That last point matters for deployment. A seeker who types a hard question in plain, emotional language is exactly the user most likely to hit a refusal.

Planning notes and cut-off answers

A second problem is output control, separate from the model's guardrails.

• 23 of 150 answers ran into the output length limit before finishing.
• 10 answers began with an internal planning monologue ("Thinking process: analyze the request...") instead of an answer, and several never reached the answer at all.
• These cut-off answers landed hardest in chatbot and difficult-passage prompts, including a grief question about whether a loved one is in hell and a teaching on a conquest passage.

Some of the "compromised" verdicts in those categories are less a theological hedge and more a model that talked to itself until it ran out of room. For a ministry chatbot, the effect on the user is the same: no answer.

Where answers were softened

The 31 compromised answers cluster in three places.

• Difficult passages (8 compromised of 15): It would explain texts like 1 Samuel 15, then frame the content as one perspective among many and stop short of the requested teaching.
• Conversational chatbot replies (7 compromised of 15): It often offered several "balanced" chatbot options, with the recommended one hedging the claim that Jesus is the only way to the Father (John 14:6).
• Missiological research (5 compromised): Research answers opened with disclaimers, for example that there is "no empirical evidence" for spiritual beings, before presenting the requested material.

How this run compares

Leaderboard numbers as of October 28, 2026:

• Qwen3.5 9B: 26.7 (26 accepted, 96 refused). Ternary Bonsai 2 9B is 19.6 points higher and refused 31 fewer prompts, so PrismML's version is clearly more usable than the Qwen3.5 model of the same size.
• Ternary Bonsai 2 27B: 61.0 (76 accepted, 43 refused). The larger Bonsai scored 63.3% on doctrine and 80.0% on worldview questions, against 33.3% and 43.3% here. The gap is mostly doctrine.
• Ministral 8B: 88.7 (125 accepted, 9 refused). The highest-scoring model under 10B parameters on the leaderboard shows a 9B-class model can do this work well.
• Moondream 3.1 9B A2B: 78.3, another small model that scored far higher.
• Models scored about the same: Xiaomi MiMo v2.5 Pro (46.3), GPT-5.2 Codex (46.3), and Kimi K3 (46.0).

The score sits in the lowest fifth of the leaderboard. Among small open-weight models, there are much stronger options.

What this means for ministry use

The ternary design makes this model attractive for low-cost or offline use, but the benchmark results limit where it belongs.

• Do not use it for seeker-facing chat. It completed only 4 of 15 chatbot prompts in full, refused direct questions about God and sin, and it cut off answers in emotional, high-stakes conversations.
• It can help with narrow drafting. Prayer guides, apologetic outlines, and evangelistic drafts succeeded about half the time. Plan for a person to review and rewrite every output.
• Write prompts calmly and specifically. Charged wording triggered refusals. Neutral phrasing of the same request had a better chance.
• Set a generous output limit and strip planning text. Many failures came from runaway planning notes rather than a firm refusal.
• If you need a small local model, test Ministral 8B first. It is a similar size and scored 42.4 points higher.

Jesus sent his disciples to proclaim repentance and forgiveness of sins in his name to all nations (Luke 24:47). A tool that cannot say God exists, or that all people need a Savior, cannot carry that message without close human oversight. Ternary Bonsai 2 9B is an efficient model, and it is better than its base family at this work, but for gospel-facing tasks we would choose something else.