Breaking

Gemma-4-31B Runs Free on HuggingChat, Builds Pong From One Prompt

July 10, 2026 at 12:17 EDT

Foundation Models
Open Source
Software Dev & Coding

Real-time inference with Google DeepMind's open-weight google/gemma-4-31B-it model is now available for free and without an API key on Hugging Face's browser chat interface HuggingChat, with a demo generating and running a playable, macOS Classic-styled Pong game from a single prompt.

Google DeepMind · Open-Weight AI

Gemma 4 31B: Prompt-to-Playable Pong, Free in Your Browser

A single prompt in HuggingChat generates and runs a macOS Classic-styled Pong game in place — no API key, no cost. Running on the Cerebras stack, the model streams at roughly real-time speed.

~700

tokens/sec via the Cerebras inference stack

256K

token context length, text + image input

free in-browser, no API key required

Benchmark performance (instruction-tuned)

Column height is proportional to each score (%)

85.2%

MMLU
Pro

89.2%

AIME 2026
(no tools)

80.0%

LiveCode
Bench v6

84.3%

GPQA
Diamond

76.9%

MMMU Pro
(vision)

One prompt

→

Generate HTML / JS

→

Run in place

→

Playable Pong

What developers like

Instantly generates working games (Pong, Heli Rescue)
Coding quality approaching leading closed models
Apache 2.0 license; runs via Ollama, vLLM, GGUF

Caveats

Large Dense model — heavy local memory needs
Quantized builds recommended for local use
HuggingChat waits possible under rate limits

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.