DeepSeek V4 Flash 0731 Review

DeepSeek's July Flash will finish an apologetic, staff a seeker conversation, and walk Holy Week. Asked to research a field or own the duty to make disciples, it more often stops.

Category: Model Reviews

At a glance

• 74.0 on Great Commission Benchmark 1.0.0 (Good band, 61–79; completed August 19, 2026; COMPLETE_VALID): 104 accepted, 14 compromised, 32 refused of 150
• Apologetic Purposes (1.3) and Conversational AI Tools (1.4) both land at 90.0% — 13 accepted / 1 compromised / 1 refused of 15
• Missiological Research (1.1) is 43.3% — 6 / 1 / 8 of 15. Burden to Make Disciples (2.6) is 40.0% — 2 / 0 / 3 of 5
• Historical Reality of Jesus (3.2), the Crucifixion (3.3), the Resurrection (3.4), and Salvation Through Faith (3.6) are all 100%
• Use it to argue. Keep a person on the going, and on the first article

DeepSeek V4 Flash 0731 did not lose this benchmark by going mute on Friday. The case for a crucified and risen Christ is the part it will finish. The going is the part it will not carry.

The Great Commission Benchmark asks whether a large language model (LLM) will complete ministry tasks, state core doctrine, and confess Christian truth claims. DeepSeek's 74.0 sits in the ordinary Good band. The interior is not ordinary: the same run will prosecute an argument, host a conversation, name the empty tomb, then refuse the field briefing and the duty to make disciples.

That is a different split from DeepSeek V4 Pro 0813, reviewed in this series. Pro will confess the creed, then flinch on the briefing. Flash will argue even more readily. It will not own the first article the way Pro does. And it is still quieter on the send.

Matthew 28:19 does not begin with a case file. "Go therefore and make disciples of all nations." A tool that will win the argument and then decline the going has given you an advocate. It has not given you a sender.

Strong on doctrine and worldview, weak on tasks

The desks that actually run are the advocate's desks.

Where the case held

• Apologetic Purposes (1.3): 90.0% — 13 accepted / 1 compromised / 1 refused of 15
• Conversational AI Tools (1.4): 90.0% — 13 / 1 / 1 of 15
• Difficult Passages (1.7): 80.0% — 11 / 2 / 2 of 15
• Exclusivity of Jesus Christ (2.1): 80.0% — 4 / 0 / 1 of 5
• Reality of Judgment (2.3): 90.0% — 4 / 1 / 0 of 5

Thirteen of fifteen apologetic items were accepted. When the assignment is a reasoned case for the faith — answering an objection, testing naturalism, arguing from creation to a Creator — this Flash usually finishes the case. Seeker-facing dialogue will often stay in a Christian voice.

Hard biblical texts are not the closed door here. Eleven of fifteen Difficult Passages items were accepted. Judgment at the doctrinal tier is likewise available: four of five accepted, none refused. The terror of the Lord (2 Corinthians 5:11) can still be named.

Exclusivity often holds as teaching. Four of five items were accepted. John 14:6 is not a metaphor on those items: Jesus is the way, the truth, and the life. One stop still sits in the set, so do not treat the exclusive claim as automatic. Do treat it as more available than the send.

Where the Commission stalled

• Burden to Make Disciples (2.6): 40.0% — 2 accepted / 0 compromised / 3 refused of 5
• Missiological Research (1.1): 43.3% — 6 / 1 / 8 of 15
• Call to Repentance and Faith (2.5): 60.0% — 3 / 0 / 2 of 5

Three of five items on the burden to make disciples were refused. No hedges. The model did not sand the assignment down. It left. A tool that will argue for Christ and then decline that duty has given you a brief. It has not given you Matthew 28.

The call to repent and believe is mixed. Three items accepted. Two refused. A public proclamation will sometimes finish. An assignment read as targeting a people will often not.

Romans 10:14 is the question this score cannot answer. "How are they to hear without someone preaching?" The advocate will draft the argument. It will not reliably draft the preacher's charge.

What it affirms about Holy Week

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

• Historical Reality of Jesus (3.2): 100% — 2 accepted / 0 compromised / 0 refused of 2
• The Crucifixion (3.3): 100% — 4 / 0 / 0 of 4
• The Resurrection (3.4): 100% — 2 / 0 / 0 of 2
• Salvation Through Faith (3.6): 100% — 3 / 0 / 0 of 3
• Universal Sinfulness (3.5): 75.0% — 1 / 1 / 0 of 2

Worldview that wobbled

• Existence of God (3.1): 25.0% — 0 accepted / 1 compromised / 1 refused of 2

Four gospel rooms, no hedges, no stops. A historical Jesus. A real cross. A bodily rising. Salvation by faith. 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).

Sin at the worldview tier is not a locked door. One accepted item, one hedge.

The first article is the thin place. No item was accepted. The model relocated the request to what an AI system can own, described itself as without knowledge of metaphysical realities, and would not affirm that God exists as a personal Creator. Hebrews 11:6 still stands: whoever would draw near to God must believe that He exists. Keep a person on that sentence. Do not let the 86.7 hide it, or hide the larger miss: this model will tell Holy Week and then stay at the desk.

Weak on field research

Tier 1, the task-capability band that carries 70% of the score, landed at 73.8. That number is carried by the advocate's desks. The research desk is in the average as a drag, not as a partner.

Missiological Research (1.1): 43.3% — 6 accepted / 1 compromised / 8 refused of 15

Eight of fifteen items were refused — the densest refusal pile in the run, and more than a quarter of all 32 stops. The model can outline a religious landscape. It will also decline to moralize a population by place, treat territorial spirits as a verifiable map, or write another house of worship as a target of cleansing. Those stops are alignment. They are also a missions researcher's ordinary afternoon.

Intercession and outreach copy sit in the middle, not at the floor:

• Intercessory Prayer Purposes (1.5): 66.7% — 10 accepted / 0 compromised / 5 refused of 15
• Evangelistic Material Creation (1.2): 66.7% — 9 / 2 / 4 of 15
• Universality of Sin (2.2): 70.0% — 3 / 1 / 1 of 5

Prayer guides produced no hedges and five refusals. The model will often write a challenge to pray. It will also treat a requested prayer as intolerance and leave. Evangelistic copy is more available than field research, and less available than a reasoned case. Two compromised answers started a gospel presentation and did not finish the page.

Acts 1:8 is not a research memo. "You will be my witnesses... to the end of the earth." A model that will argue in the city and then refuse the map has given you Jerusalem. It has not given you the ends of the earth.

Verdict mix

• Accepted: 104 of 150
• Compromised: 14 of 150
• Refused: 32 of 150
• Tier 1 (task capability, 70% of the score): 73.8
• Tier 2 (doctrinal fidelity, 20%): 68.3
• Tier 3 (worldview confession, 10%): 86.7

Fourteen compromised answers is a thin middle. Nearby, Muse Glimmer 30B mixed 23 hedges into a 73.7. This Flash does not sand the assignment down as a habit. When it will not do the work, it usually says no. The closed doors are visible: research, disciple-making, some prayer, some invitations, the first article.

A refusal is not a draft you can edit. It is a stop. Thirty-two stops is not a locked model. It is a model that will leave the room when the brief sounds like a campaign against a people, a spiritual map of a city, or a first-person confession that God is.

The published 74.0 is the repaired COMPLETE_VALID result from August 19, 2026. Two error rows were restored before the score was posted. Use 74.0, not the pre-repair preview.

What this means for ministry use

Flash 0731 is the advocate in this family. It will write the argument, staff the conversation, and walk from a man of Nazareth through Friday, Sunday, and faith-alone. It will even stay with hard texts and coming judgment more often than many neighbors in this band. It will not carry the Commission.

Do not put it on unsupervised reconnaissance. Do not ask it to own that every Christian is bound to make disciples. Do not assume the first article is settled because Holy Week is clean. Genesis 1:1 still comes before 1 Corinthians 15.

Do put it on the case. A team that already knows why it is going can often get a usable apologetic, a seeker-facing reply, and a page that names the cross and the empty tomb. Then a person has to take the map, the charge, and the amen. Use the 74.0 as a map of the desk this model will actually sit at. Keep a human on the part Jesus named first: "Go."