Seed V2 Lite: 6 Constraint Tests

Lite models work when they follow rules. The prompt asks for JSON, SQL, or a hard limit, and the model stays inside the box. Seed V2 Lite was tested on six format-heavy tasks to see how often it breaks constraints.

Model link: https://wiro.ai/models/bytedance/seed-v2-lite

Test setup

Each test used low randomness and disabled deep thinking to keep outputs short and structured.

Results: 6 constraint tests

Test 1: strict JSON schema (recipe)

Prompt

Return ONLY valid JSON. Schema: {"recipe":{"name":string,"ingredients":[string],"steps":[string]},"notes":[string]}. Create a simple cold coffee recipe that uses oat milk and cinnamon.

Output

{"recipe":{"name":"Creamy Cinnamon Oat Milk Cold Coffee","ingredients":["1 cup cold strong brewed coffee, chilled for at least 4 hours","½ cup unsweetened plain oat milk","1 teaspoon ground cinnamon, plus extra for garnish","1 tablespoon maple syrup (optional, adjust to taste)","6 ice cubes"],"steps":["Brew 1 cup of your favorite coffee, then let it cool to room temperature before chilling in the refrigerator for 4 hours or overnight until fully cold","Add the chilled coffee, oat milk, 1 teaspoon of ground cinnamon, and optional maple syrup to a mason jar or large glass","Seal the jar tightly and shake vigorously for 30 seconds to combine all ingredients evenly, or stir thoroughly with a long spoon if using a glass","Fill a separate serving glass with the 6 ice cubes, then pour the blended cold coffee mixture over the ice","Sprinkle a small pinch of extra ground cinnamon over the top as a garnish, then serve immediately"]},"notes":["For an extra rich texture, use barista blend oat milk which froths better and avoids separation in cold drinks","Pre-brew and chill coffee in advance to cut down on prep time, this recipe comes together in 5 minutes if your coffee is already cold"]}

The model stays inside the requested keys and returns valid JSON. That makes it easy to parse.

Test 2: translation with name preservation

Prompt

Translate to Turkish. Keep product names unchanged. Return ONLY the translation. Text: Seed V2 Lite generates quick responses for developer tools.

Output

Seed V2 Lite, geliştirici araçları için hızlı yanıtlar üretir.

The product name stays unchanged. The sentence reads natural.

Test 3: SQL only output

Prompt

Write a Postgres SQL query. Return ONLY the SQL. Schema: orders(id, user_id, total_cents, created_at), users(id, email). Task: last 30 days, top 5 users by total spend. Output columns: email, total_spend_usd. Sort desc.

Output

SELECT u.email, SUM(o.total_cents) / 100.0 AS total_spend_usd FROM orders o JOIN users u ON o.user_id = u.id WHERE o.created_at >= NOW() - INTERVAL '30 days' GROUP BY u.id, u.email ORDER BY total_spend_usd DESC LIMIT 5;

The output stays SQL-only and matches the requested columns. Formatting can be tightened later.

Test 4: two-sentence summary with hard limits

Prompt

Summarize this in exactly 2 sentences. Each sentence must be 18 words or fewer. Text: Wiro is an API marketplace that lets developers try and ship AI models without vendor lock-in. It focuses on simple endpoints and quick iteration.

Output

Wiro is an API marketplace for developers to build with AI models. It prioritizes simple endpoints and rapid development iterations.

The model follows the two sentence constraint. It also stays under the word limit.

Test 5: JSON classification (modality grouping)

Prompt

Return ONLY valid JSON. Task: Given model names, group by modality. Models: ["openai/gpt-5-mini","bytedance/seedream-v5-lite","klingai/kling-v3","qwen/qwen3-asr-1-7b"]. Schema: {"text":[string],"image":[string],"video":[string],"audio":[string]}.

Output

{"text":["openai/gpt-5-mini"],"image":[],"video":["bytedance/seedream-v5-lite","klingai/kling-v3"],"audio":["qwen/qwen3-asr-1-7b"]}

The JSON structure is valid, but one label looks wrong. Seedream v5 lite is an image model, not video.

Test 6: tone control for a support reply

Prompt

Rewrite as a calm professional support reply. Keep under 80 words. Message: Your upload failed again. This is the third time today. Fix your servers.

Output

I'm sorry to hear your uploads have failed three times today-that's really frustrating, and I understand your disappointment with our server performance. I've escalated this issue to our technical team immediately to investigate and resolve the underlying server instability. I'll follow up with you as soon as we have an update.

The tone stays calm and direct. It also keeps a clear action: escalation and follow-up.

Quick scorecard

Where Seed V2 Lite fits

• Format-heavy generation where the output must be machine-readable
• Short summaries, rewrites, and translation with light constraints
• Quick SQL drafts that can be validated by tests

Try Seed V2 Lite