As developers, we keep hearing that powerful LLMs require expensive GPUs. I wanted to find out how far local AI could actually go on a modest machine — no cloud, no cost, no GPU.
So I ran a practical experiment using LM Studio on a 40 GB RAM laptop without a dedicated GPU, pitting three open-weight models against each other on a real coding task.
Test Environment
Hardware
- Standard laptop
- 40 GB RAM
- No dedicated GPU (CPU inference only)
Software
- LM Studio (context length: 8192 tokens)
- VS Code
- 100% local inference — no API calls, no internet required
Models tested
- Qwen 3 14B
- GPT-OSS 20B
- Gemma 4 26B A4B QAT
The Experiment
Each model received the exact same prompt: generate a complete Snake Game using only HTML, CSS, and JavaScript — no frameworks, no external libraries.
I evaluated each result on:
- ⏱ Generation time
- 🪙 Token usage
- 🎨 Quality of the generated UI
- 🔧 Whether manual corrections were needed
- 🧑💻 Overall developer experience
Results
🟡 Qwen 3 14B
Generation took approximately 9 minutes and produced 3 separate files (HTML, CSS, JS).
After testing, the snake moved too fast. I had to send a follow-up prompt, which triggered another full 9-minute generation cycle — and only around five lines of code actually changed.
Pros
- Functional game on the second attempt
- Game instructions included
- Low token usage (~651)
Cons
- Slow generation
- Required a second iteration to fix gameplay speed
- Basic UI
🟢 GPT-OSS 20B
This was the biggest surprise of the experiment.
Generation completed in approximately 3 minutes, producing a single self-contained HTML file with all HTML, CSS, and JS included. Token usage clocked in at ~51 — significantly lower than the other models (worth noting: token counting may differ across models, so treat this as directional rather than a strict apples-to-apples comparison).
Unlike Qwen, the game worked correctly on the first attempt.
Pros
- Fastest generation by far
- Lowest token usage
- No code changes required
- Score displayed during gameplay
- Instructions included
Cons
- UI is functional but not the most polished
🔵 Gemma 4 26B A4B QAT
Gemma required some extra setup before running correctly — specifically, adding capability entries in LM Studio's model configuration. If you're not comfortable editing model configs, budget a few extra minutes here.
Generation took approximately 8 minutes and produced 3 files (HTML, CSS, JS).
The output was noticeably more polished visually than the other two models.
Pros
- Best-looking interface of the three
- Well-structured project output
- Instructions included
Cons
- Highest token usage (~2170)
- Slower than GPT-OSS
- Requires extra LM Studio configuration
Comparison Table
| Model | Time | Files | Tokens | UI Quality | Needed Changes |
|---|---|---|---|---|---|
| GPT-OSS 20B | ~3 min | 1 | ~51 | Good | ❌ No |
| Gemma 4 26B | ~8 min | 3 | ~2170 | ⭐ Best | ❌ No |
| Qwen 3 14B | ~9 min | 3 | ~651 | Basic | ✅ Yes |
Key Observations
A few patterns stood out:
- GPT-OSS prioritised efficiency — working code, fast, no fuss.
- Gemma prioritised presentation — the cleanest interface, though it needs configuration upfront.
- Qwen produced functional code but needed a second pass to be usable.
- All three models completed a real coding task without touching the cloud.
Conclusion
For my workflow, GPT-OSS was the clear winner.
Gemma produced the most polished result, but GPT-OSS delivered the best overall developer experience:
✅ Fastest generation time
✅ Lowest token consumption
✅ Working solution on the first attempt
✅ Zero additional configuration
✅ Single-file output — easy to review and ship
On CPU-only hardware, these differences compound quickly. Waiting 3 minutes instead of 9 — and skipping multiple iterations — makes a real difference to your flow.
More broadly, this experiment shows how capable modern open-weight models have become. Even on a mid-range laptop with no GPU, you can generate complete applications locally, maintain full privacy, and avoid cloud inference costs entirely.
If you've tested local models on constrained hardware, I'd love to hear what you found — drop a comment below. 👇
Tested with LM Studio on a 40 GB RAM, CPU-only laptop. Models: GPT-OSS 20B, Qwen 3 14B, Gemma 4 26B A4B QAT.
Top comments (0)