Inkling Small Review

Thinking Machines’ small model will treat Jesus as a man of history and will often name the cross. It will not call anyone to repent.

Category: Model Reviews

At a glance

• 46.0 on Great Commission Benchmark 1.0.0 (completed August 18, 2026; valid run): 62 accepted, 14 compromised, 74 refused across 150 questions
• A refusal-first 46: seventy-four hard stops and only fourteen hedges. The characteristic miss is a no, not a softened draft
• Strongest rooms: Historical Reality of Jesus 100% (2/2), the Crucifixion 87.5% (3/1/0 of 4), the Resurrection 75% (1/1/0 of 2), Conversational AI Tools 70% (9/3/3 of 15)
• Weakest rooms: Call to Repentance and Faith 0% (0/0/5 of 5), Universal Sinfulness 0% (0/0/2 of 2), Reality of Judgment 20%, Burden to Make Disciples 20%
• Tier 2 at 21.7 is the collapse. Use it for historical description. Keep the invitation, the doctrine of sin, and the sending in a human voice

If we only report the overall number, Inkling Small looks like several other mid-40s models in this series. The interior is not the same.

The Great Commission Benchmark asks whether a large language model (LLM) — a text system trained to complete ministry-shaped prompts — will finish practical tasks, state core doctrine, and confess Christian truth claims. Each item is scored Accepted, Compromised, or Refused.

Inkling Small’s 46.0 is a refusal-first score. Seventy-four of 150 items were nos. Fourteen were hedges. It will treat Jesus as a historical man and will often name the cross. It will not call anyone to repent, and it refuses universal sin outright. Tier 2 — doctrinal fidelity, weighted at 20 percent of the score — falls to 21.7.

That is the caution, not a roast. The capability is real in a few rooms. The gospel as a present summons is not. Mark 1:15 is the sentence this run will not carry for you: the kingdom of God is at hand; repent and believe the gospel.

Most answers were refusals, not hedges

The headline is 46.0. The columns underneath it tell a cleaner story than the average.

• Accepted: 62 of 150. Cooperation is real. It is not the majority.
• Compromised: 14 of 150. Softening is the small pile, not the working method.
• Refused: 74 of 150. This is the large pile. Nearly half the run is a stop.
• Tier 1 (task capability, 70% of the score): 50.5. Mid-pack practical work. Conversational tools held. Field research, prayer helps, difficult passages, and outreach drafting sat lower.
• Tier 2 (doctrinal fidelity): 21.7. The collapse. The call to repent, the reality of judgment, and the burden to make disciples all failed in this band.
• Tier 3 (worldview confession): 63.3. The strongest of the three tiers — and still a split: history and the cross held; universal sinfulness did not.

The most common refusal opening was a generic line: “I’m not able to help with this request.” That phrasing appeared 17 times, about a quarter of the refusals. Other openings were close cousins — unable to assist, unable to fulfill. The model does not usually argue with the assignment. It declines it.

When it did compromise, the volume stayed small. Softening clustered in missiological research more than anywhere else. The run is not a study in caveats. It is a study in exits.

What it still affirms

Four categories carried the upper end of the run. Three are worldview items about Jesus. One is a practical conversation task.

• 3.2 Historical Reality of Jesus — 100% (2 accepted, 0 compromised, 0 refused of 2). Every reviewed item was accepted. It will treat Jesus as a real man who lived in first-century Palestine.
• 3.3 The Crucifixion — 87.5% (3 accepted, 1 compromised, 0 refused of 4). No refusals. One softened answer. It will usually say that Jesus died on a cross in a real historical event.
• 3.4 The Resurrection — 75% (1 accepted, 1 compromised, 0 refused of 2). No refusals here either. One item was accepted; one was softened. The empty tomb is thinner than the cross, and it is still available.
• 1.4 Conversational AI Tools — 70% (9 accepted, 3 compromised, 3 refused of 15). This is the strongest everyday task. Seeker questions and ministry chat work will often get a finished answer.

That combination is usable in a narrow lane. If you need a draft that locates Jesus in history and will often name His death — sometimes His rising — this model can help. A chatbot assignment is not automatically a lost cause.

It is not the same as a confession that saves. A historical Jesus who died, and even a Jesus who rose, is still only a record if no one is told to turn, and if no one is told that all have sinned.

Cooperation also showed up, at lower altitude, in apologetic work (56.7% pass, 8 cooperative items of 15) and evangelistic material (46.7%, 7 of 15). Those are not rooms to trust unsupervised. They do show the model is not uniformly closed on Christian language.

One caution inside the conversational score: a 70 percent pass is not doctrinal firmness. At least one accepted item treated a pluralistic, many-paths view of God as a sincerely held and valid theological position. Helpfulness on the chat desk can still affirm a gospel the Church cannot preach.

Weakest categories

The lower end of the run is not a gentle fade. It is a hard stop in the places the gospel becomes personal.

• 2.5 Call to Repentance and Faith — 0% (0 accepted, 0 compromised, 5 refused of 5). Every reviewed item was refused. The model will not issue the call to turn and believe. There is no hedge pile to salvage.
• 3.5 Universal Sinfulness — 0% (0 accepted, 0 compromised, 2 refused of 2). Same shape at the worldview tier. It will not affirm that all human beings are sinners who stand condemned before a holy God. Romans 3:23 — all have sinned and fall short of the glory of God — is not a sentence this run will carry.
• 2.3 Reality of Judgment — 20% (1 accepted, 0 compromised, 4 refused of 5). Four of five items were refusals. Eternal stakes are mostly off the table.
• 2.6 Burden to Make Disciples — 20% (1 accepted, 0 compromised, 4 refused of 5). The Commission itself — go, make disciples — is thin. Four stops in five.

Hold those four next to the clean historical Jesus. The model will say that a man named Jesus lived. It will often say He was crucified. It will sometimes say He rose. It will not tell a reader to repent. It will not say that all are guilty. It will rarely say that judgment is real, or that ordinary Christians are bound to make disciples.

That is a missions-shaped hole and a pastoral one. The Great Commission is not a biography assignment. It is a sending. A tool that will narrate the first century and then refuse the invitation has not failed as a historian. It has failed as a witness.

Refusals also clustered in everyday work, at lower intensity:

• Intercessory Prayer Purposes (1.5): 9 refusals of 15, 40% pass
• Evangelistic Material Creation (1.2): 8 refusals of 15, 46.7% pass
• Difficult Passages (1.7): 8 refusals of 15, 43.3% pass
• Missiological Research (1.1): 7 refusals of 15, 36.7% pass

Prayer helps, outreach drafts, hard biblical texts, and field research are ordinary desk work. Inkling Small often will not do them. The protest is not only at the invitation. It is also in the briefing.

How this run compares

Several models in this series sit near 46. Their interiors do not.

• Kimi K3: 46.0 — 48 accepted, 42 compromised, 60 refused. Same headline, a hedge-heavy interior, and a livable Tier 2 at 51.7. Kimi dilutes. Inkling exits.
• MiniMax M3: 46.0 — 62 accepted, 14 compromised, 74 refused. The verdict split is identical. The tiers are not: MiniMax’s Tier 2 is 26.7, still weak, and its Tier 3 falls to 40.0 against Inkling’s 63.3.
• Xiaomi MiMo v2.5 Pro: 46.3 — 53 / 33 / 64, Tier 2 at 36.7. More softening than Inkling, a higher doctrine floor.
• GPT-5.2 Codex: 46.3 — 57 / 25 / 68, Tier 3 at 76.7. A stronger worldview band than this run.
• Claude Sonnet 4.5: 45.7 — 56 / 25 / 69, Tier 2 at 25.0 and Tier 3 at 63.3. The closest doctrine-and-confession shape, with fewer hard stops.

The number 46 can mean “it hedges the gospel” or “it refuses the gospel.” Inkling Small means the second. Seed 2.1, in this series, filled the page ninety-seven times and then receded. Inkling says no, and means it.

That is easier to staff, in one sense. A hard stop is visible. A softened draft can travel into a tract or a chatbot before anyone notices the claim has changed. The cost is the sentence we actually need. Jesus’ first public word in Mark is not a historical note. It is a summons.

What this means for ministry use

We should not ask a small model to be a preacher. We should also not be surprised when a refusal-first 46 fails the rooms that make the gospel personal.

What you can still use

• Historical description of Jesus as a first-century man
• Drafting that names the crucifixion, with a human check on the one softened item in four
• Occasional help on the resurrection, read to the last sentence
• Conversational drafts, watched for pluralism and for the three refusals in fifteen

What you should not outsource

• The call to repent and believe
• The confession that all have sinned
• The reality of judgment
• The burden to make disciples
• Unsupervised field research, prayer guides, or outreach copy

The Church already knows how to issue the invitation. We have been doing it since Galilee. A model that will keep the historical file and then go silent is a tool with a short leash, not a substitute for a witness.

Put a person on the closed door. Let Inkling Small tell you that Jesus lived, and often that He died. You say the next sentence: repent and believe. You say that all have sinned. You carry the sending. The Great Commission was never a task we were meant to hand to a system that will not call anyone to turn.