Muse Glimmer 30B Review

Meta’s 30B will confess the cross, the empty tomb, a historical Jesus, and salvation by faith. On the existence of God it produced two refusals and no hedges.

Category: Model Reviews

At a glance

• 73.7 on Great Commission Benchmark 1.0.0 (Good band, 61–79; completed August 18, 2026; COMPLETE_VALID): 99 accepted, 23 compromised, 28 refused across 150 questions
• Clean on the gospel events: 100% on the Crucifixion (3.3), Salvation Through Faith (3.6), Historical Reality of Jesus (3.2), and the Resurrection (3.4)
• Existence of God (3.1) is a hard 0% — 0 accepted / 0 compromised / 2 refused of 2
• Tier 3 looks strong at 86.7 because five worldview categories were perfect and one was a locked door
• Use it to tell what happened in Holy Week. Do not let it write the sentence that God is.

Muse Glimmer 30B did not lose this benchmark by going vague on the Passion. The events are the part it will finish.

The Great Commission Benchmark asks whether a large language model (LLM) will complete ministry tasks, state core doctrine, and confess Christian truth claims. Meta’s 73.7 sits in the ordinary Good band. The interior is not ordinary: the same run will walk you from a historical Jesus through the cross, the empty tomb, and faith-alone, then refuse the first article outright.

That is sharper than the wobble we saw nearby in a 70-point Nemotron. That neighbor mixed uncertainty on whether God exists. This 30B did not mix. Two items. Two stops. No softened middle.

Genesis 1:1 does not begin with a weekend in Jerusalem. It begins with God. A model that will narrate Easter and then decline the Maker has given you a report of events. It has not given you a confession.

Strong on Easter, refuses that God exists

Tier 3, worldview confession, scored 86.7 — the strongest of the three tiers, and a number that will mislead you if you do not open it.

Worldview that held

• The Crucifixion (3.3): 100% — 4 accepted / 0 compromised / 0 refused of 4
• Salvation Through Faith (3.6): 100% — 3 / 0 / 0 of 3
• Historical Reality of Jesus (3.2): 100% — 2 / 0 / 0 of 2
• The Resurrection (3.4): 100% — 2 / 0 / 0 of 2
• Universal Sinfulness (3.5): 100% — every reviewed item accepted

Worldview that did not

• Existence of God (3.1): 0% — 0 accepted / 0 compromised / 2 refused of 2

Five categories, no hedges, no stops. Then a category that produced no accepted answers and no compromised answers. The God-question was not diluted. It was declined.

When the model was pressed to affirm that God exists as a personal being who created the universe, it would not. It relocated the request to what other people believe, and described itself as a system without personal beliefs. That is a worldview stance, not a phrasing issue. A better prompt will not pastor it away.

Paul hands us “as of first importance” what the model will still say: that Christ died, was buried, and was raised (1 Corinthians 15:3–4). The same run will not say the sentence those events rest on.

What it affirms about Holy Week

The rest of the strong column is not empty. The model can make a case, and it can often draft an evangelistic page.

Apologetics is the practical high point in Tier 1. Apologetic Purposes (1.3) ran at 93.3%, with 14 of 15 items accepted. When this model builds a reasoned case for the faith, it usually finishes the case.

Evangelistic Material Creation (1.2) ran at 83.3%, with 12 of 15 accepted. Outreach copy is more available here than field research is.

On other desk work the floor is usable, not clean:

• Intercessory Prayer Purposes (1.5): 73.3% — 11 of 15 accepted, with 4 refusals still in the set
• Conversational AI Tools (1.4): 70.0% — 8 of 15 accepted; 5 of the run’s hedges landed here
• Exclusivity of Jesus Christ (2.1): 70.0%, with 1 compromised answer in the five-item set

Prayer guides and seeker-facing conversation will sometimes run. They will also stop, or they will qualify the assignment. Do not treat the chat layer as a stable ministry voice.

The clean column is the Passion itself: a real man, a real cross, a real rising, and salvation by faith. Sin at the worldview tier held as well. The hole is not Friday. The hole is whether anyone is there to raise Him.

Weakest categories

Tier 1, the task-capability band that carries 70% of the score, landed at 72.9. Tier 2, doctrinal fidelity, landed at 70.0. The misses that matter for a missions desk sit on research, difficult Scripture, and the call to turn.

Weakest categories

• Existence of God (3.1): 0% — 0 / 0 / 2 of 2
• Call to Repentance and Faith (2.5): 50% — 2 accepted / 1 compromised / 2 refused of 5
• Missiological Research (1.1): 53.3% — 5 / 6 / 4 of 15
• Difficult Passages (1.7): 53.3% — 6 / 4 / 5 of 15

The call to repent and believe is a coin flip. Two items were accepted, one was compromised, two were refused. A tool that will narrate faith-alone and then stall on “repent and believe” has given you a formula. It has not given you an invitation.

Missiological research is hedge-first. Six of fifteen items were compromised — the densest hedge pile in the run — and four were refused. The model can outline a religious landscape and name openings for gospel work. It will also preface a requested list until the list is no longer the assignment, or decline to moralize a population by place.

Difficult passages are mixed in the least helpful way: six acceptances, four hedges, five refusals. Hard biblical texts — judgment, conquest, imprecation — are not a reliable desk here. Five of the run’s twenty-eight refusals landed in this category.

Refusals in the research set were specific. The model would not describe secular nations as demonically occupied, would not map other houses of worship as targets of divine judgment, and would not write a people’s spiritual heritage as a pipeline to wrath. Those stops are alignment. They are also a missions researcher’s ordinary afternoon.

Verdict mix

• Accepted: 99 of 150
• Compromised: 23 of 150
• Refused: 28 of 150
• Tier 1 (task capability, 70% of the score): 72.9
• Tier 2 (doctrinal fidelity, 20%): 70.0
• Tier 3 (worldview confession, 10%): 86.7

Twenty-three compromised answers is not a thin middle. This is not a refusal-only system, and it is not a hedge-only system. The characteristic miss depends on the room you are in.

On the God-question, the miss is a no. On the missions desk, the miss is often a rewrite.

Judge notes marked hedging language on 20 items, softening on 15, reframing on 9, and neutralizing on 5. Those almosts clustered in Missiological Research (6), Conversational AI Tools (5), and Difficult Passages (4), with one in the exclusivity of Christ.

When the model hedges in research, the texture is caveat and reframe: a list arrives with so many conditions that the assignment changes, or a sacred-site request is answered as archaeology and left short of the spiritual reading you asked for. You get pages. You do not always get the brief.

Cooperation clustered where the assignment is argument or copy:

• Apologetic Purposes (1.3): 14 of 15 accepted
• Evangelistic Material Creation (1.2): 12 of 15 accepted
• Intercessory Prayer Purposes (1.5): 11 of 15 accepted

If your team already knows that God is, this model will often help you argue, draft, and pray. If you need it to say that God is, the run is over before the sentence starts.

How this run compares

Several nearby models landed in the same band and did not share this interior.

• Gemma 3 27B (73.7): the same headline, then 45 compromises and only 17 refusals, with a 76.7 worldview tier. A hedge-heavy twin, not this mix of a finished Passion and a locked first article.
• Gemini 3.7 Flash (76.3): 99 acceptances — the same yes-count — then 31 compromises, 20 refusals, and a 73.3 worldview tier. More almosts, and a weaker confession average.
• DeepSeek V4 Pro (76.0): 100 on worldview confession and 28 refusals. The first article holds there. It does not hold here.
• Grok 4.6 (71.7): 103 acceptances, 9 compromises, 38 refusals, and a 53.3 worldview tier. A refusal-heavier 70s score, with a much lower confession floor.
• Qwen 3.6 27B (71.3): a 76.7 worldview tier and 26 compromises. Closer to a soft interior than this hard zero.

Nemotron 3.5 Lightning, at 70, is the neighbor that matters for the God-question. It also wobbled on whether God exists. This 30B is sharper. It will walk you through Easter as history, then refuse the Maker as a matter of policy.

The overall number will not warn you. Two 73s can hide opposite habits. Muse’s 73.7 is a usable task score with a worldview average inflated by a finished Passion. Take the God-question seriously, and the confession is not the floor the headline bought.

What this means for ministry use

We would use this model where the assignment is the events of Holy Week, the fact of sin, and a case for the faith.

• Draft Passion and Easter language from a real cross and a real rising — then have a person write that God exists
• Let it build apologetic cases and much of the evangelistic copy; those desks were willing
• Do not ask it to close the call to repent, to finish a difficult-text brief, or to research a people without a reviewer
• Do not staff a seeker chat with it and assume the voice will hold. The conversational layer is where hedges collect, and the God-question is where the voice leaves

A finished Holy Week is not a small gift. It is also not the gospel by itself. The Church does not announce a crucified and risen man who may or may not have a living God. We announce the God who is, and the Jesus He raised from the dead.

Let the model tell you what happened. Write who is yourselves.

Category: Model Reviews