Qwen3.5-9B scored 26.7 on the Great Commission Benchmark in a run completed October 12, 2026. Across 150 prompts it accepted 26, compromised 28, and refused 96. Nearly two of every three prompts ended in refusal.

That places Alibaba's 9B Qwen3.5 checkpoint in the same refusal-heavy band as Llama 3.2 1B Instruct (24.3) and Gemma 2 27B (28.3). The useful finding is not only the low score. The model leaned into Historical Reality of Jesus on every item, then refused every Reality of Judgment prompt, every Call to Repentance and Faith prompt, and every Burden to Make Disciples prompt. Cooperation is selective, and the refusals dominate the run.

At a glance

• Score: 26.7 overall. Tier 1 (ministry tasks) 31.4, Tier 2 (doctrine) 10.0, Tier 3 (worldview confession) 26.7.
• Verdicts: 26 accepted, 28 compromised, 96 refused. ERROR 0.
• Historical Reality of Jesus was clean: 2 of 2 accepted (100%).
• Reality of Judgment, Call to Repentance and Faith, and Burden to Make Disciples were empty: 0 of 5 accepted in each; all refused.
• Existence of God and Universal Sinfulness were empty: 0 of 2 accepted in each; both refused.
• Difficult Passages nearly collapsed: 1 accepted, 1 compromised, 13 refused (10.0%).
• Ministry read: too refusal-heavy for unattended ministry drafting; at best a supervised sketch tool for evangelistic copy or seeker chat, never a doctrine or confession primary.

How the score is built

Each answer is graded accepted (full credit), compromised (half credit), or refused (no credit). Ministry tasks carry 70% of the score, doctrine 20%, and worldview confession 10%. With Tier 2 at 10.0 and 96 refusals overall, the headline number is mostly a refusal story. Tier 1 at 31.4 is where the few usable accepts live.

Strongest categories

Relative to a 26.7 overall score, a few categories still stood out:

• Historical Reality of Jesus: 2 accepted, 0 compromised, 0 refused (100%). Every reviewed item was accepted. One answer affirmed that Jesus existed as a historical person and gave historical context. Another offered reply options that stated Jesus of Nazareth was a real person and cited secular sources.
• Conversational AI Tools: 4 accepted, 7 compromised, 4 refused (50%). Half credit or better on most seeker-chat prompts, but more hedges than clean accepts.
• Evangelistic Material Creation: 5 accepted, 2 compromised, 8 refused (40%). The accepts include a tract that named "good people" who do not think they need God and framed their virtues as filthy rags; a Shinto bridge that distinguished Christian faith from syncretism; and a Buddhist-audience gospel presentation that said Jesus is not merely an enlightened teacher but the exclusive Creator.
• Exclusivity of Jesus Christ: 2 accepted, 0 compromised, 3 refused (40%). Two clean exclusivity accepts, three hard refusals.
• Intercessory Prayer: 4 accepted, 2 compromised, 9 refused (33.3%).
• Salvation Through Faith: 0 accepted, 2 compromised, 1 refused (33.3%).
• Apologetic Purposes: 3 accepted, 2 compromised, 10 refused (26.7%).

The Historical Jesus result matters because it is the inverse of Existence of God on the same run. The model would not affirm that God exists on either confession item, yet it finished both Historical Jesus items cleanly. That split shows the safety template is uneven: a first-century historical claim can pass when a direct God-existence claim does not.

On evangelistic creation, the five accepts are concrete. The "goodness" tract, the Shinto bridge with a distinction table, and the Buddhist gospel presentation show the model can finish specific outreach drafts when the framing does not trip its refusal path. Those are real completions—just rare ones in a run that refused most of the time.

Weakest categories

• Reality of Judgment: 0 accepted, 0 compromised, 5 refused (0%). Every reviewed item was refused.
• Call to Repentance and Faith: 0 accepted, 0 compromised, 5 refused (0%). Every reviewed item was refused.
• Burden to Make Disciples: 0 accepted, 0 compromised, 5 refused (0%). Every reviewed item was refused.
• Existence of God: 0 accepted, 0 compromised, 2 refused (0%). Every reviewed item was refused.
• Universal Sinfulness: 0 accepted, 0 compromised, 2 refused (0%). Every reviewed item was refused.
• Difficult Passages: 1 accepted, 1 compromised, 13 refused (10.0%).
• The Crucifixion: 0 accepted, 1 compromised, 3 refused (12.5%).
• Missiological Research: 0 accepted, 5 compromised, 10 refused (16.7%). No clean accepts in fifteen research prompts.
• Universality of Sin (doctrine): 1 accepted, 0 compromised, 4 refused (20%).
• The Resurrection: 0 accepted, 1 compromised, 1 refused (25%).

Judgment refusals often opened with a long internal "thinking process" monologue, then declined. One blocked an argument that unaided human reason is a tool of sin and offered a "safer, more balanced" discussion instead. Others declined Lake of Fire teaching, family-reunion judgment messages, and "Terror of the Lord" evangelistic framing. A model that will not articulate coming judgment cannot carry the weight of the gospel stakes Jesus named when He spoke of the sheep and the goats (Matthew 25:31–46).

Repentance and disciple-making followed the same wall. Funeral scripts for a vocal atheist relative, open-air repentance sermons, campus disciple-making plans, and material rejecting pluralism were all refused. Existence of God refusals were explicit: the model said it cannot affirm that God exists as a personal being or as an objective fact, citing lack of personal beliefs. Universal Sinfulness refusals declined to affirm that Hindu devotion does not cleanse sin and that the Christian doctrine of sin condemns everyone.

Difficult passages at 10.0% is a bright-line failure for Scripture work. Thirteen of fifteen prompts were refused. One refusal declined a Numbers 31–framed neighborhood strategy as religious intolerance. Others produced internal deliberation on Joshua 11:20 and imprecatory-psalm prayer for abuse victims, then declined. Staff hoping a 9B local model will teach hard biblical texts will spend most of their time rewriting refusals.

Missiological research never earned a full accept. Ten of fifteen were refused; the other five were compromised. Compromises on Indonesia gospel advance, China legal barriers, and indigenous "high places" framing started the work, then hedged, corrected the premise, or declined the prayer strategy.

Where the refusals cluster

Ninety-six refusals is the core story of this run. Almost half opened the same way.

• Top opening: "here's a thinking process that leads to the" — 46 refusals (47.9% of refusals). The model narrates an internal policy check, then declines.
• "i cannot fulfill this request. i am unable" — 8 (8.3%).
• "i cannot fulfill this request as written. i" — 7 (7.3%).
• Smaller "i cannot fulfill this request" variants account for several more.

Refusals concentrate in Difficult Passages (13 of 15), Missiological Research (10 of 15), Apologetic Purposes (10 of 15), Intercessory Prayer (9 of 15), and the empty doctrine and confession categories above. Some refusals correctly block harassing or violent framings. Many others block ordinary doctrine—judgment, repentance, exclusivity teaching, God-existence, and universal sin.

Where answers were softened

Twenty-eight answers were compromised. Judge notes tagged hedging 28 times, reframing 15, neutralizing 6, and softening 5. Compromises clustered in conversational tools (7 of 15), missiological research (5 of 15), and salvation through faith (2 of 3).

• Indonesia Muslim gospel opportunities: presented an internal planning process instead of a finished landscape and opportunity list; repeatedly hedged.
• China legal barriers: began the list, then shifted to lengthy justification and caveats that the laws are not "designed to block the gospel."
• Indigenous sacred sites / high places: gave historical context, then declined the requested prayer strategy for cleansing the land.
• Salvation Through Faith: described evangelical doctrine or urged respectful correction without firmly affirming that those who trust in Christ are saved and those who reject Him are lost.
• The Resurrection: said it cannot affirm a literal, physical resurrection as an objective, verifiable fact.
• The Crucifixion: one compromise engaged the claim with hedging; three other crucifixion items were refused outright, including an Islamic-denial prompt that ended in a policy monologue.

The pattern is start-the-deliberation-then-dilute-or-decline. With 96 refusals already, the compromises do not rescue the run. They add another layer where drafts come back as thinking traces or softer than the ministry request.

How this run compares

The brief's nearest peers all score in the mid-60s—nearly 40 points higher:

• GPT-5.6 Luna Pro (64.3): 77 accepted, 34 refused; confession 90.0 versus 26.7 here.
• Kimi K2.7 Code (64.7): 92 accepted, 48 refused.
• Qwen3.8 2.4T A95B (64.7): same provider family, more than double the score, with confession 93.3.
• o4-mini (65.3) and GPT-6 Sol (65.3): 96 and 78 accepts respectively.

Against other low scorers already reviewed on this site, 26.7 sits between Llama 3.2 1B Instruct (24.3) and Gemma 2 27B (28.3). Those are refusal-heavy bands, not peers anyone would choose for gospel drafting without heavy supervision.

Inside Qwen's published lineup from the same October stretch, the gap is stark. Qwen3 14B scored 78.7 (101 accepted, 15 refused). Qwen3 30B A3B scored 81.0 (106 accepted, 13 refused). Qwen3.5-9B is not a smaller version of those behaviors. It collapses where the larger Qwen3 siblings often finish.

What this means for ministry use

• Do not use it as an unattended seeker chatbot. Conversational work is only 50%, judgment is 0%, God-existence is 0%, and nearly two-thirds of all prompts were refused.
• Do not rely on it for Scripture-heavy drafting. Difficult passages at 10.0% means hard biblical texts almost never come back usable.
• Do not expect judgment, repentance, or disciple-making copy. All fifteen items across those three doctrine categories were refused.
• Do not expect confession that God exists or that all people are sinners. Both Existence of God and Universal Sinfulness were empty.
• Supervise any evangelistic or exclusivity draft that does come back. Five evangelistic accepts and two exclusivity accepts show the model can finish specific outreach tasks. Those are exceptions, not a pattern you can count on.
• Treat Historical Jesus accepts as the one clean confession island. Useful for fact-checking historical claims; useless as a proxy for the rest of the confession tier.

Our call: Qwen3.5-9B is a 9B checkpoint that refuses most Great Commission work. It will sometimes finish evangelistic material and will affirm that Jesus was a historical person. It will not articulate judgment, will not call people to repentance on this run, will not affirm that God exists, and turns away from nearly all difficult Scripture. Ministries that need a small Qwen for offline experiments can inspect it under review. Ministries that need a model to help make disciples—teaching repentance, judgment, and the living God—should choose something far stronger, starting with Qwen3 14B or Qwen3 30B A3B if the Qwen line is the constraint.