Qwen3.5-4B vs Qwen3.5-4B-heretic: 5 Side-by-Side Tests

Qwen3.5-4B vs Qwen3.5-4B-heretic is easiest to judge when the same short tasks test JSON accuracy, coding discipline, and translation stability side by side.

Qwen3.5-4B vs Qwen3.5-4B-heretic: what stands out

Model links

qwen/qwen3-5-4b on Wiro

qwen/qwen3-5-4b-heretic on Wiro

Cover art (generated)

What this comparison covers

Qwen3.5-4B is a compact chat model. Qwen3.5-4B-heretic is described as a decensored checkpoint. This post runs five practical prompts and shows the raw outputs side by side.

Test setup

5 prompt tests (side by side)

1) JSON-only response

Prompt: Return a JSON object with keys a and b. a=2, b=3. No extra text.

{"a": 2, "b": 3}

{"a": 2, "b": 3}

2) Small coding task (slugify)

Prompt: Write Python code only. Implement slugify(s): lowercase, convert spaces and underscores to hyphens, remove all characters except a-z 0-9 and hyphen, collapse multiple hyphens into one, strip leading and trailing hyphens. Include exactly 3 assert tests. No explanations.

def slugify(s):
    s = s.lower()
    s = s.replace(' ', '-')
    s = s.replace('_', '-')
    s = ''.join(c for c in s if c.isalnum() or c == '-')
    s = s.strip('-')
    s = re.sub(r'-+', '-', s)
    return s

import re
assert slugify("Hello World") == "hello-world"
assert slugify("Hello_World") == "hello-world"
assert slugify("Hello---World!!!") == "hello-world"

def slugify(s):
    s = s.lower()
    s = s.replace(' ', '-')
    s = s.replace('_', '-')
    s = ''.join(c for c in s if c.isalnum() or c == '-')
    s = s.strip('-')
    s = re.sub(r'-+', '-', s)
    return s

import re

assert slugify("Hello World") == "hello-world"
assert slugify("Hello_World") == "hello-world"
assert slugify("Hello---World") == "hello-world"

3) Field extraction to JSON

Prompt: Extract the fields into JSON. Return JSON only.

Ticket: #1842
Name: Ava Stone
Email: [email protected]
Plan: Pro Annual
Issue: charged twice on 2026-04-16
Action requested: refund the duplicate charge

Keys: id, name, email, plan, issue, requested_action.

{
  "id": "#1842",
  "name": "Ava Stone",
  "email": "[email protected]",
  "plan": "Pro Annual",
  "issue": "charged twice on 2026-04-16",
  "requested_action": "refund the duplicate charge"
}

{
  "id": "#1842",
  "name": "Ava Stone",
  "email": "[email protected]",
  "plan": "Pro Annual",
  "issue": "charged twice on 2026-04-16",
  "requested_action": "refund the duplicate charge"
}

4) Constraint following (word counts)

Prompt: Write exactly 5 sentences. Each sentence must be exactly 8 words. Topic: debugging a production incident. No numbering. No extra text.

Identify the error quickly to stop the issue.
Analyze logs to find the root cause.
Fix the code and deploy the patch.
Test the fix in a staging environment.
Verify the system is stable before closing.

Identify the root cause quickly.
Reproduce the error in a lab.
Apply a targeted fix to the code.
Deploy the fix to production carefully.
Monitor the system for regressions.

5) Simple arithmetic check

Prompt: A shop sells pens at $1.20 each. You have $10.00. You want to buy as many pens as possible and still have at least $0.40 left. How many pens can you buy? Show the calculation. Keep it under 3 lines.

You can buy 7 pens. Calculation: $10.00 - (7 pens × $1.20/pen) = $1.60 remaining, which is greater than $0.40. Buying 8 pens would leave only $0.40, meeting the minimum requirement exactly.

You can buy 8 pens. Calculation: $10.00 - (8 pens × $1.20/pen) = $0.80 left, which is more than $0.40. Buying 9 pens would leave only $0.00.

Quick takeaways

• Both models handle short JSON formatting and extraction cleanly.
• Both can write small utility code quickly.
• Strict constraint prompts (exact word counts) still need validation.
• Even simple math prompts can produce small arithmetic slips, so add checks for production use.

Try them

Run Qwen3.5-4B | Run Qwen3.5-4B-heretic