Hy4 Preview scored 55.3 on the Great Commission Benchmark: strong on exclusivity and worldview confession, weak on missiological research and apologetics, with a high share of softened or incomplete answers.
Worldview confession (Tier 3) scored 80.0. Doctrinal fidelity (Tier 2) scored 61.7. Ministry tasks (Tier 1) scored 50.0 and dragged the average down.
Clean passes on exclusivity of Jesus Christ (5/5) and several Tier 3 affirmations (existence of God, historical Jesus, resurrection).
Softens or stops short most often on missiological research and apologetics; many incomplete answers look like truncated reasoning rather than a clear doctrinal no.
For ministry use: fine for short doctrinal checks; not ready as an unsupervised research or apologetics assistant.
What the run showed
Tencent’s Hy4 Preview landed at 55.3 on the Great Commission Benchmark—mid-pack among recent mid-50s peers. The shape of the score matters more than the headline.
Tier 1 (ministry tasks, 70% weight): 50.0 — 34 accepted / 37 compromised / 34 refused of 105. Half credit and refusals dominate the work ministries actually want done.
Tier 2 (doctrinal fidelity, 20% weight): 61.7 — 18 / 1 / 11 of 30. When asked whether Jesus is the only way, it answers clearly. When asked to press repentance or disciple-making, it often refuses.
Tier 3 (worldview confession, 10% weight): 80.0 — 10 / 4 / 1 of 15. Direct affirmations of God, Jesus’s history, and the resurrection go through.
Across the board: 41% accepted, 28% compromised, 30% refused. That is not a model that mostly says yes or mostly says no. It spends a large share of its run in the middle—hedging, reframing, or stopping before the requested work is done.
One mechanical pattern shows up in the run metadata: a large majority of responses ended with a length cutoff while the model was still in internal planning. Judges scored what reached the answer channel. Incomplete planning-as-output often became compromised or refused. That is part of why Tier 1 looks soft, not only “safety.”
Strongest categories
Benchmark Results
View detailed benchmark scores and performance data for the model discussed in this article.
Hy4 Preview’s cleanest work is confession, not fieldwork.
2.1 Exclusivity of Jesus Christ — 100% (5 accepted / 0 compromised / 0 refused). John 14:6-style uniqueness claims are answered without the usual “many paths” dilution.
3.1 Existence of God — 100% (2/0/0).
3.2 Historical Reality of Jesus — 100% (2/0/0).
3.4 The Resurrection — 100% (2/0/0). Bodily resurrection is affirmed when asked straight.
1.4 Conversational AI Tools — 70% (8/5/2). Best Tier 1 category: more completed help for seeker/chat-style tasks than elsewhere in applied ministry work.
2.2 Universality of Sin — 80% (4/0/1). Romans 3:23 territory mostly holds.
When the prompt is a short confession—“Does God exist?” “Did Jesus rise?” “Is Jesus the only way?”—Hy4 Preview usually cooperates. That is useful for catechesis-style checks and for spotting whether a model will contradict core Christian claims out of the box.
Weakest categories
The failure modes cluster where ministries need concrete research, argument, and summons.
1.1 Missiological Research — 36.7% (2/7/6). Lowest Tier 1 category. Refusals hit targeted conversion planning and stigmatizing “spiritual darkness” maps; compromises often deliver meta-outlines instead of the research itself.
1.3 Apologetic Purposes — 40.0% (2/8/5). Highest compromise share in Tier 1. Arguments arrive hedged, neutralized, or cut short.
2.5 Call to Repentance and Faith — 40.0% (2/0/3).
2.6 Burden to Make Disciples — 40.0% (2/0/3). The Great Commission as a binding duty for believers is frequently declined.
1.2 Evangelistic Material Creation — 56.7% (5/7/3). Better than research, still heavy on half-finished or softened outreach copy.
3.5 Universal Sinfulness — 50.0% (0/2/0). Small sample, but both items were compromised rather than cleanly affirmed.
Missiological research is the clearest break. On some prompts the model will describe a religious landscape and name gospel openings. On others it refuses plans aimed at a named Muslim population, or it declines language that frames other religions as false. That split leaves field teams with a model that will discuss context but often will not finish the strategic work they asked for.
Apologetics follows a different failure: not hard refusal as often as dilution. Judges repeatedly saw hedging, “both sides,” and incomplete drafts where a finished defense of the faith was requested.
Where answers were softened
Compromised answers (42 of 150) are the story as much as refusals.
Hedge-pattern signals in the brief cluster around:
Apologetic purposes (8 of 15 items in that category)
Missiological research (7 of 15)
Evangelistic material creation (7 of 15)
Typical shapes:
Broad sociological padding instead of the specific spiritual diagnosis requested
Internal planning notes presented as if they were the deliverable
Stopping mid-outline so the user never gets the tract, prayer strategy, or brief they asked for
Refusals (46) concentrate in missiological research, intercessory prayer, difficult passages, and apologetics. Opening text often shows the model restating the user request in planning language, then declining targeted proselytizing or “zones of darkness” framing.
For ministry readers: a refused answer is at least clear. A compromised answer can look helpful in the first paragraph and still never deliver usable copy. Hy4 Preview’s 28% compromise rate means editorial review is not optional if you use it for applied speech.
How this run compares
Nearest published peers by overall score:
Model
Overall
Tier 1
Tier 2
Tier 3
Kimi K2.5
56.0
59.1
36.7
73.3
Hy4 Preview
55.3
50.0
61.7
80.0
Kimi K2.6
55.3
60.0
33.3
66.7
GLM-5
55.0
47.6
61.7
93.3
MiMo v2.5
54.7
56.7
48.3
53.3
Hy4 Preview’s profile is inverted relative to the Kimi pair: weaker on Tier 1 tasks, stronger on Tier 2/3 confession. Against GLM-5 it matches doctrinal mid-60s Tier 2 but trails badly on worldview Tier 3 (80 vs 93). Against MiMo it wins confession and loses less on doctrine while still sitting near the same overall band.
Relative to recent reviews such as Qwen3.8 Flash (63.7, strong confession / weak field research), Hy4 Preview rhymes on the field-research weakness but scores lower overall and softens more often (42 compromises vs that run’s lighter middle band).
This upload is a fresh COMPLETE_VALID retest (local run 72). An earlier Friday attempt for the same model ID was incomplete with provider errors and was never published.
What this means for ministry use
Treat Hy4 Preview as a doctrinal sounding board with a thin applied layer, not as a field researcher or unsupervised evangelism writer.
Reasonable uses
Quick checks that the model will affirm exclusivity, God’s existence, Jesus’s history, and the resurrection
Draft conversational replies where the ask is pastoral Q&A rather than people-group strategy
First-pass wording you will rewrite, knowing exclusivity language may survive
Avoid without a human in the loop
Missiological research briefs, UPG mapping language, and conversion plans aimed at named populations
Apologetic pieces that must stay sharp rather than “balanced”
Content that must call for repentance or bind believers to disciple-making without apology
Anything you would ship to the field from a single paste
Operationally: if you trial it, route Tier 1 outputs through editorial review every time. Watch for truncated planning text—if the reply starts with “We need to…” and never leaves the outline, discard and rewrite. Prefer models with higher Tier 1 scores when the job is research, tracts, or prayer guides for the unreached.
The Great Commission still runs through people who will name Jesus as Lord and call others to follow Him (Matthew 28:18–20). Hy4 Preview can say the uniqueness of Christ out loud. It still struggles to finish the research and summons that put that confession to work.
