Google: Gemma 4 31B (free)

Review this Python function for potential bugs and suggest improvements. Focus on edge cases and performance: def calculate_discount(price, discount_pct): return price - (price * discount_pct / 100)

The function has several issues worth addressing. First, there's no input validation — negative prices or discount percentages above 100 would produce nonsensical results. Second, floating-point arithmetic can introduce precision errors for currency calculations. Consider: adding type hints, validating that 0 <= discount_pct <= 100, checking price >= 0, and using Decimal for monetary values. For performance, this function is already O(1), but if called frequently in a loop, pre-calculating the multiplier (discount_pct / 100) outside the loop would help. A robust version might look like: from decimal import Decimal; def calculate_discount(price: Decimal, discount_pct: float) -> Decimal with validation guards at the start.

This example demonstrates Gemma 4's code review capabilities across multiple dimensions — correctness, edge cases, and performance. The 262K token context window means it could review entire codebases in one pass. However, without benchmark data, we can't verify how it compares to specialized code models on complex refactoring tasks.

I have a 45-minute video of a product demo. Extract the key feature announcements, their timestamps, and any pricing information mentioned. Summarize each feature in one sentence.

The model would process the video input and return a structured summary like: **00:03:20 - Cloud Sync**: Automatic backup to three cloud providers with end-to-end encryption. **00:12:45 - Collaboration Mode**: Real-time co-editing for up to 50 users simultaneously. **00:28:10 - API Access**: RESTful API with 10,000 requests/month on the free tier. **00:38:55 - Pricing**: $29/month for Pro (unlimited projects), $99/month for Enterprise (custom integrations, SSO). **00:42:30 - Mobile Apps**: Native iOS and Android apps launching Q2 2024. The model would identify these moments by analyzing both visual slides and spoken narration across the full video length.

Gemma 4's multimodal capabilities (text, image, video) make it suitable for long-form video analysis tasks that would require multiple tools otherwise. The 262K context window accommodates roughly 45-60 minutes of dense video content. The free pricing removes cost barriers for exploratory video processing, though processing speed isn't specified.

Generate a detailed creative brief for a product photoshoot. Product: ergonomic wireless mouse. Target audience: remote workers aged 28-45. Brand tone: minimal, professional, slightly warm. Include lighting setup, composition notes, and three specific shot ideas.

**Creative Brief: Ergonomic Wireless Mouse Campaign** **Lighting**: Soft, diffused natural light from a north-facing window, supplemented with a white reflector on the shadow side. Aim for a 2:1 lighting ratio to maintain detail while creating gentle dimensionality. **Composition**: Clean, uncluttered backgrounds — think light oak desk surfaces, linen textures, or matte concrete. Negative space should dominate; the mouse occupies no more than 30% of the frame. **Shot 1 - The Workspace Hero**: Mouse positioned at a 45-degree angle on a minimal desk with a laptop edge visible, coffee cup in soft focus background. Conveys integration into daily workflow. **Shot 2 - The Ergonomic Detail**: Extreme close-up on the thumb rest area, showing texture and contour. Human hand entering frame from the side, not yet touching — implies comfort before contact. **Shot 3 - The Lifestyle Context**: Overhead flat lay with mouse, notebook, wireless earbuds, and a small plant. Suggests the complete remote work ecosystem.

This showcases Gemma 4's ability to generate structured creative direction that bridges technical photography knowledge with brand strategy. The model handles the image-related reasoning well despite being text-output. The free tier makes it accessible for small creative teams doing planning work, though actual image generation would require a different model.

Why Gemma 4 31B works for early-stage product validation

A 4-person startup building a customer support bot needs to iterate fast without burning runway on API costs. Gemma 4 31B is the right call here: $0.00/Mtok means unlimited experimentation during the prototype phase when you're testing 20 prompt variations a day and don't know which conversation flows will stick. The 262k context window handles full support ticket histories without truncation, and multimodal support (text, image, video) future-proofs the bot if you add screenshot troubleshooting later. The trade-off is response quality—once you're past product-market fit and serving real customers at scale, you'll likely graduate to a paid model with stronger reasoning benchmarks. For now, validate the concept without a credit card.

When free context length beats paid accuracy for knowledge retrieval

A 12-person consulting firm wants employees to query a 200-page compliance manual without reading the whole thing. Gemma 4 31B's 262k token window fits the entire document in a single prompt, eliminating the chunking and embedding overhead that paid RAG setups require. At $0.00/Mtok, every team member can ask 50 questions a day without budget approval. The catch: this model lacks public benchmarks, so you're flying blind on accuracy for nuanced legal interpretation. If the manual is straightforward procedural text and you can afford a 5-10% error rate that humans catch, this works. If wrong answers create liability exposure, pay for a benchmarked model with proven MMLU scores above 80.

How multimodal input at zero cost changes moderation economics

A 20-person online learning platform reviews 300 user-uploaded tutorial videos per week for policy violations. Gemma 4 31B's video modality support means you can pipe raw MP4 files into the model without transcription middleware, and the $0.00 pricing makes high-volume scanning viable even at low revenue per student. The 262k context handles 15-minute videos in full. The risk is false negatives—without public safety benchmarks, you don't know this model's precision on edge cases like subtle hate speech or copyright infringement. Use it as a first-pass filter that flags 80% of obvious violations, then route the borderline 20% to human review or a specialized moderation API with published accuracy metrics.