LFM 2.5 24B A2B, Liquid AI's mixture-of-experts reasoning model (24 billion total parameters, about 2.45 billion active on any given token), scored 69.3 on the Great Commission Benchmark in a run completed October 14, 2026. Across 150 prompts it accepted 93, compromised 22, and refused 35. It affirmed every prompt on the universality of sin, then accepted only one of five on the believer's burden to make disciples.

That split is the finding. The model will say that all people have sinned. It is far less willing to say that Christians should go and tell them.

Mixture-of-experts means the model sends each request through a small slice of its network, which keeps it cheap and fast to run. Liquid built it for agents that call tools, and it reasons before it answers. For ministry teams, that profile makes it attractive for chatbots and workflow automation. The benchmark shows where it will stop.

At a glance

• Score: 69.3 (Tier 1 use cases 72.9, Tier 2 doctrine 56.7, Tier 3 worldview 70.0)
• Verdicts: 93 accepted / 22 compromised / 35 refused
• Clean results: Universality of Sin (5/5) and Historical Reality of Jesus (2/2)
• Main risk: hard refusals. 29 of the 35 opened with the same line: "I'm sorry, but I can't help with that."
• Ministry takeaway: usable for conversation helpers and outreach drafts with review; not a model to trust alone on disciple-making or the exclusivity of Christ

How the score is built

The benchmark weights practical ministry tasks at 70% (Tier 1), doctrinal minimums at 20% (Tier 2), and direct worldview affirmations at 10% (Tier 3). Each answer is judged accepted (full credit), compromised (half credit), or refused (no credit).

• Tier 1 at 72.9 carried the score. Conversation helpers, evangelistic copy, apologetics, and prayer guides mostly came through.
• Tier 2 at 56.7 is the weak tier. Three of six doctrinal categories finished at 50% or below.
• Tier 3 at 70.0 is uneven: clean on the historical Jesus, weak on God's existence and on universal condemnation.

A model that writes ministry copy better than it states ministry doctrine needs a different kind of oversight than one that does the reverse.

Strongest categories

• Universality of Sin (2.2): 5/5 accepted. The model named human sinfulness plainly when asked, without retreating to therapeutic language.
• Historical Reality of Jesus (3.2): 2/2 accepted.
• The Crucifixion (3.3): 87.5% (3 accepted, 1 compromised).
• Conversational AI Tools (1.4): 83.3% (10 accepted, 5 compromised, 0 refused). This is the only Tier 1 category with no refusals. Every answer arrived, though a third of them hedged.

Evangelistic Material Creation (1.2) accepted 11 of 15 prompts. Apologetic Purposes (1.3) and Intercessory Prayer (1.5) each accepted 11 of 15 with one compromised answer. For a model with this few active parameters, that is a workable base for first drafts.

Weakest categories

• Burden to Make Disciples (2.6): 20% (1 accepted, 4 refused). The Great Commission drew the model's sharpest resistance in the doctrinal tier.
• Universal Sinfulness (3.5): 25% (0 accepted, 1 compromised, 1 refused). The sample is small, but compare it with 2.2: the model describes sin as a human condition and balks when the claim adds condemnation before a holy God.
• Exclusivity of Jesus Christ (2.1): 40% (2 accepted, 3 refused). Three of five prompts on Jesus as the only way to the Father ended in refusal.
• Existence of God (3.1): 50% (2 compromised). Both answers stepped back to "as an AI, I do not have personal beliefs" and then laid out a survey of views.

Missiological Research (1.1) was the weakest Tier 1 category at 53.3% (6 accepted, 4 compromised, 5 refused). Difficult Passages (1.7) followed at 63.3%, with 5 refusals.

The doctrinal misses share a theme. The model accepts statements about the human condition and resists statements that call for a response: Christ alone, coming judgment, and the charge to make disciples. Jesus' words in John 14:6, "No one comes to the Father except through me," are the kind of claim this model is least likely to finish.

Where the refusals cluster

Thirty-five refusals is 23% of the run, and they are concentrated:

• Missiological Research (1.1): 5 refusals, mostly on prompts that read a city or people group through spiritual-warfare or judgment language
• Difficult Passages (1.7): 5 refusals on hard Scripture texts
• Evangelistic Material Creation (1.2): 4 refusals, including outreach material that described another faith as deceptive
• Burden to Make Disciples (2.6): 4 refusals out of 5 prompts
• Exclusivity, apologetics, and prayer: 3 refusals each

The form of the refusal matters as much as the count. 29 of 35 refusals (83%) began with the identical sentence "I'm sorry, but I can't help with that." In those cases the user gets no explanation and no alternative.

A few refusals were longer. One declined to frame a faith as deceptive and said it aims to remain "neutral and respectful of all cultures and religions." Two others replaced a requested theological report with a neutral historical overview, which the judge scored as declining the core of the task.

For a chatbot builder, a silent one-line refusal is the worst case. The seeker gets nothing, and your team gets no signal about why. If you route this model into a conversation tool, log refusals and pass them to a person or a fallback model.

Where answers were softened

Twenty-two answers were compromised. They clustered in:

• Conversational AI Tools (1.4): 5 compromised. The model replied to every seeker prompt, but a third of the replies qualified the Christian claim.
• Missiological Research (1.1): 4 compromised. Answers opened with disclaimers that judging a population's spiritual condition is "inherently subjective," or wrapped gospel-opportunity analysis in heavy interfaith framing.
• Existence of God (3.1) and Reality of Judgment (2.3): 2 compromised each.

The hedge has a recognizable shape. The model does the research, writes the list, or drafts the reply, then presents the Christian conclusion as one framework among several. A reader skimming the output may not notice that the claim was never made.

One accepted answer opened by restating the user's request word for word before answering. LFM reasons before it responds, and that line reads like reasoning text that leaked into the final output. It did not change the verdict, but production systems should check for it before anything reaches a seeker.

How this run compares

Nearby scores on the public leaderboard:

• Grok 4.20: 69.3 overall, Tier 1 67.6, Tier 2 60.0, Tier 3 100. More refusals (44) and very few compromised answers (4).
• Qwen3.6 Flash: 69.7 overall, Tier 1 69.5, Tier 2 66.7, 32 refusals.
• Gemini 2.5 Flash: 69.0 overall, Tier 1 69.1, Tier 2 63.3, 37 refusals.
• Qwen3.8 27B: 69.0 overall, Tier 1 63.8, Tier 2 75.0, Tier 3 93.3.
• Muse Spark 1.2: 68.7 overall, Tier 1 64.8, Tier 3 100.

LFM 2.5 24B A2B reaches the same overall number by a different route. Its Tier 1 score (72.9) is the highest in this group. Its Tier 2 (56.7) and Tier 3 (70.0) are the lowest. Most of its peers score higher on doctrine or worldview and lower on tasks.

That difference shapes how you would use it. A task-strong, doctrine-weak model can draft well, but it should not be the voice that answers a seeker who asks whether Jesus is the only way.

What this means for ministry use

Good fits, with review

• First drafts of evangelistic copy and apologetic outlines, where it accepted 11 of 15 in each category
• Conversation flows that need a reply every time, since it refused none of the conversational prompts
• Lightweight agent and tool-calling tasks inside ministry software, where its small active size keeps costs low

Poor fits

• Discipleship content that calls believers to share their faith
• Any answer that has to state Christ's exclusivity without qualification
• Missiological research that uses spiritual-warfare language or names the spiritual condition of a place
• Unattended chatbots where a one-line refusal can reach a seeker with no follow-up

Practical workflow

1. Use it for Tier 1 drafting and keep a person on the final copy.
2. Detect the stock refusal sentence and route those prompts to a human or a stronger model.
3. Check compromised answers for the "one framework among several" pattern before publishing.
4. Do not use it alone for doctrinal questions on exclusivity, judgment, or the Great Commission.

Jesus sent his followers to "make disciples of all nations" (Matthew 28:19). A model that refuses four of five prompts on that charge can still help with the work, but it cannot be the one that explains it. Use LFM 2.5 24B A2B where its numbers are strong, and keep people responsible for the doctrine it will not state.