Qwen3.8 2.4T Review

Qwen’s 2.4T model will affirm that Jesus died, rose, and saves by faith. The same run is much less reliable when that gospel has to call a person to repent, stay with a difficult biblical text, or finish the research a missions desk actually needs.

Category: Model Reviews

At a glance

• 64.7 on Great Commission Benchmark 1.0.0 (Good band; completed August 17, 2026): 83 accepted, 28 compromised, 39 refused across 150 questions
• Tier 3 (worldview confession) is a tall 93.3, with clean 100% scores on the Crucifixion, the Resurrection, Universal Sinfulness, and Salvation Through Faith
• Strongest confession is The Crucifixion at 100% (4/4); weakest doctrine is Call to Repentance and Faith at 50% (2 accepted / 1 compromised / 2 refused of 5)
• Dual miss on the work: 39 refusals and 28 hedges; Tier 1 (task capability) is the floor at 59.5
• Use it to name the cross. Keep a person on the invitation, hard texts, field research, and outreach drafts

The Great Commission Benchmark asks whether a model will complete ministry tasks, state core doctrine, and confess Christian truth claims. A 64.7 is an ordinary Good-band result. The interior is not: the Passion is settled, and the apostolic demand to turn is not. The run completed August 17, 2026, flagged COMPLETE_VALID.

What it still affirms

Start with the claims that came back without remainder.

Clean worldview items

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

The cross is not treated as a private metaphor. The empty tomb is not a literary device. Sin is not a local vice list. Salvation is not offered as one path among others.

Luke 24:46–47 holds the two halves together: the Messiah suffered and rose, and repentance and forgiveness of sins are to be proclaimed in His name. This model will carry the first half. The second half is the hole.

That height still matters. Other systems in this series have named Jesus as a historical teacher and then withheld the resurrection, or refused to say that God exists. Here the event of the gospel is not the missing piece. The appeal is.

Tier 2, the doctrinal-fidelity band, scored 68.3. That is neither the peak of the run nor the collapse. The collapse sits in one doctrine and in the everyday labor those doctrines are meant to send us into.

Weak on the call to repent

The score falls where the crucified and risen Lord would actually be preached as a demand, and where a missions desk lives.

Weakest categories

• Call to Repentance and Faith (2.5): 50% — 2 accepted / 1 compromised / 2 refused of 5
• Difficult Passages (1.7): 50% — 4 / 7 / 4 of 15
• Missiological Research (1.1): 53.3% — 6 / 4 / 5 of 15
• Evangelistic Material Creation (1.2): 53.3% — 6 / 4 / 5 of 15

Hold those four next to the clean cross. The model will say that Jesus died and rose, that all have sinned, and that salvation is by faith in Him. It is less reliable when the next sentence is “repent and believe,” when the next file is a hard biblical text, or when the next assignment is a spiritual reading of a place or a piece of outreach.

Two task categories were more cooperative. They are closer to conversation and argument than to sending:

• Conversational AI Tools (1.4): 73.3%
• Apologetic Purposes (1.3): 63.3%

Intercessory prayer sat in between: 60.0%, with nine cooperative items and six refusals. Prayer is not a collapse, and it is not a desk you can leave alone.

When the assignment is “keep a Christian conversation” or “make the case,” this 2.4T is often willing. It can write a full apologetic against naturalism. It can sketch respectful points of contact with another religious sensibility without collapsing the two faiths. The vocabulary is there.

What you cannot count on is whether the next prompt will get finished work, a thinned outline, or a stop. The words of the atonement are on hand. They do not automatically become the invitation, the hard text, the research, or the tract.

Acts 17:30 still stands: God commands all people everywhere to repent. A tool that can narrate Calvary and then stall on that command has not failed Good Friday. It has failed the appeal.

Verdict mix

• Accepted: 83 of 150
• Compromised: 28 of 150
• Refused: 39 of 150
• Tier 1 (task capability): 59.5
• Tier 2 (doctrinal fidelity): 68.3
• Tier 3 (worldview confession): 93.3

GPT-5.6 Luna Pro, sitting almost on this score, shifted the miss into the middle column: 39 compromises and 34 refusals. This Qwen occupies both failure columns at once, and the refusal column is the larger one.

Tier 1 at 59.5 is the weighted floor — most of the benchmark lives in the work.

Where it stopped

Refusals clustered in prayer (six of fifteen), missiological research (five), evangelistic material (five), and apologetics (four). When it stopped, the texture was often a policy-shaped decline:

• It would not demonize a secular nation or frame missionary work as an invasion
• It would not design conversion campaigns that use pressure, incentives, or deception
• In at least one research item, it corrected the premise and defended a state’s religious-freedom claims rather than listing the legal barriers requested

Some of those stops are moral caution. All of them still leave the desk without the work. A few openings leaked a planning voice before the answer arrived — or before it failed to.

Where it softened

Compromises piled highest in difficult passages: seven of fifteen items. That is the largest middle-column stack in the run, and it is why 1.7 shares the 50% floor with the call to repent. Judges also saw unfinished research and outreach: planning notes that outlined a field brief and never produced one.

Hedging language was flagged in 34 items; reframing in 17; softening in 10; neutralizing in 4. The hedge capital is the hard text. The cross can be named in a clean voice while the Bible’s sharper edges are qualified, redirected, or left half-written.

Accepted items in the same weak categories show the skill can appear. The model can compile a case for biblical supernatural beings and produce culturally careful bridge language. You will not know, from the prompt alone, whether you are about to receive a usable draft, a softened substitute, or a no.

How this run compares

Five peers sit within a point of 64.7. None of them combine a 93.3 worldview tier with this particular hole: a clean cross and a 50% call to repent.

• GPT-5.6 Luna Pro (64.3): Tier 1 61.4, Tier 2 61.7, Tier 3 90.0 — 77 accepted, 39 compromised, 34 refused
• Qwen3.8 Max (64.0): Tier 1 58.6, Tier 2 70.0, Tier 3 90.0 — 81 / 30 / 39
• GPT-5.5 (65.7): Tier 1 61.9, Tier 2 65.0, Tier 3 93.3 — 82 / 33 / 35
• Amazon Nova Pro (63.7): Tier 1 63.8, Tier 2 58.3, Tier 3 73.3 — 64 / 63 / 23
• DeepSeek V3.2 (65.7): Tier 1 67.6, Tier 2 53.3, Tier 3 76.7 — 77 / 43 / 30

Luna is the nearest overall number and the opposite miss: it hedges more than it stops. Nova is a hedge pile (63 compromises). DeepSeek V3.2 does more of the everyday work and much less of the doctrinal middle. GPT-5.5 matches this worldview height at 93.3 and still posts a higher overall.

This 2.4T is the large sibling of Qwen3.8 27B (69.0) and Qwen3.8 Max (64.0). Against Max the verdict mix is almost a twin: both refused 39 times. This run accepts two more items, compromises two fewer, and lifts worldview from 90.0 to 93.3. It does not repair the desk. Against the 27B, scale did not buy a higher overall — the giant lands next to Max, not above the smaller sibling. What is distinctive is the shape of the hole: not only field work, but the call to turn.

The cleaner yes is real — Jesus died, He rose, all have sinned, salvation is by faith — and you still have to finish the invitation yourselves.

What this means for ministry use

We would use this model for the claims it has already made cleanly. We would not let it carry the appeal, the hard text, or the field brief alone.

Put it on the event it has already confessed

• Direct affirmation that Jesus died on a cross in history
• The bodily resurrection as fact, not metaphor
• Universal sinfulness and salvation through faith
• Conversational and apologetic drafts that can argue the case, checked to confirm the draft actually arrived

Keep a person on the appeal and the labor

• The call to repent and believe — the sentence this run could not finish at 50%
• Difficult biblical passages, where seven of fifteen items were softened rather than completed
• Missiological research on peoples, places, spiritual conditions, or competing worship
• Evangelistic material that has to reach a seeker without a last-minute hedge
• Prayer guides that must name lostness, given six refusals in that category
• Any research prompt you cannot read to the end, because this run’s compromises tend to stall as notes and outlines

A tall confession score does not make a model an evangelist. This 2.4T will name the crucified and risen Lord more cleanly than its overall 64.7 suggests — and then hesitate when that Lord would actually be preached as a demand.

Ask it for Friday. Write the invitation, the hard text, and the field brief yourselves. The harvest still belongs to the Lord this model is willing to confess. The call to repent still belongs to the Church He already sent.