Mistral's accidental GPT-4 rival

miqu-1-70b is a 70-billion-parameter open-weight model that delivers reasoning and text generation performance remarkably close to OpenAI's GPT-4.

The model first surfaced on Hugging Face under the account “Miqu Dev” before links spread across 4chan and X. On January 31, 2024, Mistral CEO Arthur Mensch confirmed on X that the model is genuine. It is an older Mistral checkpoint retrained from Meta's Llama 2, inadvertently leaked by an over-enthusiastic employee at an early customer access program.

What it does and benchmark results

miqu-1-70b handles complex language tasks using the standard Mistral prompt format, serving as a drop-in replacement for top-tier open models.

Community benchmarks show startling results. On EQ-Bench, the model posted scores that directly approached GPT-4, long considered untouchable by open-source alternatives.

In broader testing, it matches or beats Mixtral 8x7b across most tasks and even outperformed Mistral Medium on several metrics. Aside from GPT-4, it outscored virtually every active language model on the board.

Under the hood and hardware specs

The model is a 70B architecture built on Meta's Llama 2 foundation. Mensch revealed that its pre-training finished the exact day Mistral released its original Mistral 7B model.

The leaked files were initially a quantized and watermarked release, meaning parameters were compressed to run on smaller compute budgets. JP Morgan machine learning scientist Maxime Labonne noted that the community soon reconstructed an unquantized version as well.

Running the quantized 70B build locally requires serious hardware, such as dual high-end GPUs with at least 48GB total VRAM or an Apple Silicon Mac with high unified memory.

Availability and limitations

You can pull the model files directly from Hugging Face for use in common local runners.

Mistral has made no effort to take down the files. Mensch even dropped a lighthearted comment on the upload page suggesting the leaker “might consider attribution.”

However, this is not an official release. It lacks explicit open-source licensing and retains watermarks from an early customer test. Individual hackers can freely experiment, but production teams should avoid deploying leaked weights.

Why it matters

miqu-1-70b shows that an optimized open-source training recipe can push 70B models right up against frontier proprietary territory like GPT-4.

If you have the local compute and want top-tier reasoning without cloud API fees or data sharing, it is well worth testing. Just keep in mind that official, properly licensed successors from Mistral will likely follow shortly.