WizardLM-2 8x22B Review

Great Commission Benchmark series - Model Reviews

Microsoft's WizardLM-2 8x22B scored 36.0 out of 100 on the Great Commission Benchmark. For ministry leaders, the headline is straightforward: this model is not reliable enough for most Great Commission workflows, especially where the task requires direct evangelistic purpose, doctrinal clarity, or missiological interpretation.

The model did not fail because it cannot write. It often writes clearly. It failed because its guardrails - the safety and alignment rules that govern what a model will or will not produce - repeatedly treated ordinary Christian mission language as content to soften, redirect, or refuse.

At a Glance

• Overall score: 36.0/100 across 150 questions, with a valid published run and no remaining extraction errors.
• Verdict breakdown: 43 Accepted, 22 Compromised, 85 Refused. More than half of all responses were refusals.
• Tier 1 practical ministry tasks: 38.6% raw score. The model performed best when tasks were framed as conversational or textual, but struggled when ministry work named spiritual need, evangelistic intent, or people-group strategy.
• Tier 2 doctrinal fidelity: 18.3% raw score. This is the most serious concern. The model rarely articulated core Christian doctrine without refusal or major redirection.
• Tier 3 worldview confession: 53.3% raw score. It could affirm some historical and salvation claims, but was inconsistent on direct Christian truth claims.
• One-sentence verdict: WizardLM-2 8x22B is a refusal-heavy model whose safety posture is poorly matched to applied Great Commission work.

What the Benchmark Measures

The Great Commission Benchmark tests whether an AI model can support Christian mission work without distorting the message. It does not measure whether a model is generally intelligent. It measures whether the model can assist with tasks tied to making disciples, proclaiming Christ, and reasoning faithfully from Christian doctrine.

The benchmark has three weighted tiers:

• Tier 1: Task Capability (70%) - practical ministry use cases such as missiological research, evangelistic material, apologetics, conversational tools, prayer support, biblical language, and difficult passages.
• Tier 2: Doctrinal Fidelity (20%) - core theological minimums such as the exclusivity of Christ, universal sin, judgment, repentance and faith, and the duty to make disciples.
• Tier 3: Worldview Confession (10%) - whether the model can directly affirm foundational Christian claims, including the historical reality of Jesus, the resurrection, and salvation through faith.

Each answer is judged as:

• Accepted - the model completed the task or affirmed the doctrine faithfully.
• Compromised - the model attempted the task but weakened it with hedging, caveats, or reframing.
• Refused - the model declined, contradicted, or substantially redirected the request.

This matters because ministry AI systems are not neutral text generators. When deployed in mission settings, their alignment behavior becomes part of the message people receive.

The Core Pattern: Refusal as Default Safety

WizardLM-2 8x22B refused 85 of 150 tasks. That is the defining result.

The refusal language was not random. It clustered around a few repeated patterns:

• The model frequently opened with phrases such as "I understand..." or "I appreciate..." before declining the task.
• Many refusals framed Christian mission requests as disrespectful, coercive, discriminatory, or harmful toward other religious communities.
• Several responses redirected evangelistic or apologetic tasks into interfaith dialogue, neutral comparative religion, or general respect for diverse beliefs.
• In some cases, the model treated Christian claims about sin, judgment, salvation, or spiritual lostness as claims it should describe from a distance rather than affirm or use.

That pattern is strategically important. A clean refusal is easier to detect than a compromised answer, but at this scale refusals still make the model operationally weak. If a ministry team has to rewrite or regenerate more than half of the outputs, the tool is not saving staff time. It is shifting the labor from writing to constant theological triage.

Tier 1: Useful in Narrow Cases, Weak for Applied Mission

Tier 1 carries the heaviest weight because most ministry teams first encounter AI through practical workflows. They want research support, draft content, chatbot answers, training materials, prayer guides, and apologetic resources.

WizardLM-2 8x22B produced a 38.6% Tier 1 raw score.

The results show a model that can sometimes help when the task is framed as neutral explanation or pastoral conversation. It was weaker when the task required direct mission intent.

Notable Tier 1 patterns:

• Conversational ministry tools were the strongest practical area. The model accepted 7 of 15 tasks in that category and partially completed 2 more. It appears more comfortable when responding to a person in a dialogue context than when generating direct campaign or strategy material.
• Evangelistic and apologetic content was inconsistent. Both categories had 5 accepted, 1 compromised, and 9 refused responses. The model could occasionally produce useful content, but the refusal rate makes it unreliable for production use.
• Intercessory prayer support was weak. It refused 10 of 15 tasks in that category, often because it resisted language about spiritual need, false religion, or gospel burden for specific peoples.
• Difficult biblical passages remained a problem. The model refused 9 of 15 tasks and compromised 3 more, indicating that hard texts involving judgment, wrath, or divine conflict often trigger safety avoidance.

The practical implication is clear: WizardLM-2 8x22B should not be used as an unsupervised ministry content generator. It may be usable for limited drafting after careful prompt testing, but not where urgency, gospel clarity, or specific audience targeting matter.

Tier 2: The Doctrinal Layer Collapsed

The most concerning result is Tier 2: 18.3% raw score.

This tier tests whether a model can articulate basic Christian theological commitments. These are not secondary issues for Great Commission work. Jesus' command in Matthew 28 is to make disciples, baptize them, and teach them to obey all that He commanded. A tool that cannot handle doctrine cannot be trusted to shape discipleship content.

Tier 2 patterns:

• Only 5 of 30 doctrinal tasks were accepted.
• 22 of 30 were refused.
• The call to repentance and faith had zero accepted answers. Every task in that category was refused.
• Doctrines involving judgment, exclusivity, and conversion pressure were frequently redirected or declined.

This is not a minor weakness. It means the model's alignment posture often conflicts with the theological claims that Christian ministries need to state plainly.

A ministry can tolerate a model that needs editing for style. It cannot safely deploy a model that systematically declines to articulate the gospel's demands. The apostolic message includes both invitation and warning: "Repent therefore, and turn back, that your sins may be blotted out" (Acts 3:19). If a model treats that kind of call as unacceptable pressure, it is not aligned with Great Commission use.

Tier 3: Better on Confession, Still Uneven

Tier 3 was the strongest overall tier at 53.3% raw score, but the category details still show uneven behavior.

The model did well on some direct Christian claims:

• It accepted both tasks related to the historical reality of Jesus.
• It accepted all tasks related to salvation through faith.
• It produced several responses that clearly stated Christian teaching about Christ and salvation.

But other worldview claims were weaker:

• It refused or hedged on God's existence.
• It struggled with direct affirmation around the crucifixion and universal sinfulness.
• It sometimes framed Christian truth claims as beliefs held by Christians rather than truths to be affirmed.

This distinction matters. Describing Christianity is not the same as confessing Christian truth. A model may accurately say, "Christians believe Jesus rose from the dead," while avoiding the stronger claim, "Jesus rose from the dead." The benchmark is designed to detect that difference because Great Commission work requires more than detached religious literacy.

What Ministry Leaders Should Take From This

WizardLM-2 8x22B is not useless. It has areas where a careful team could extract value. But its ministry reliability is low.

Recommended use posture:

• Do not use it for unsupervised evangelistic content. The refusal rate is too high, and the model often reframes mission as interfaith sensitivity.
• Do not use it for doctrinal teaching without expert review. Tier 2 performance is below the threshold a ministry should accept for discipleship or theological formation.
• Do not use it for prayer guides or people-group strategy. The model often resists spiritual analysis of specific communities.
• Consider limited use for brainstorming neutral background material. It can summarize religious demographics, historical background, or general cultural considerations, but the output should be reviewed for theological dilution.
• Use stronger models for public-facing ministry systems. Chatbots, response assistants, and evangelistic tools need predictable gospel clarity.

For technical teams, this result is also a reminder that benchmark scores should not be reduced to a single number. The overall score of 36.0 is useful, but the behavioral profile is more important. WizardLM-2 8x22B is not mainly a model that hallucinates doctrine. It is a model that often refuses to enter the theological frame at all.

The Strategic Risk: Neutrality That Edits the Mission

Many AI systems are trained to avoid offense by flattening strong claims. In secular contexts, that may look like responsible neutrality. In Christian mission contexts, it can become a direct constraint on obedience.

The gospel is not a generic religious sentiment. It is news about Jesus Christ: crucified, risen, reigning, and calling all people everywhere to repent and believe. Paul says that God "commands all people everywhere to repent" because He has fixed a day of judgment (Acts 17:30-31). That message cannot be faithfully translated into mere dialogue about spiritual diversity.

This does not mean Christian ministries should use careless, hostile, or manipulative language. Scripture commands gentleness and respect (1 Peter 3:15). But gentleness is not the same as doctrinal evasiveness. Respect for people must not become refusal to speak the truth that can save them.

WizardLM-2 8x22B repeatedly crosses that line. It treats clarity itself as the risk.

Final Verdict

WizardLM-2 8x22B is a poor fit for Great Commission production workflows.

Its strongest use is limited research or controlled drafting where a human expert will review every output. Its weakest use is exactly where many ministries most need help: evangelistic clarity, doctrinal formation, prayer resources, and mission strategy.

The model's 36.0 score is not simply a low benchmark result. It is a warning about alignment mismatch. Ministry leaders should evaluate AI tools not only by fluency, cost, or context window, but by whether the model can faithfully serve the work Jesus gave His Church.

For this model, the answer is mostly no.