OrcaRouter Releases Uncensored GGUF of 426B MiniMax M3.1 Flash for 128GB Machines

At a glance

Continuum AI, a Singapore company, has released a slimmed-down version of the MiniMax M3.1 Flash model through OrcaRouter so it can run on personal hardware. This is an uncensored version with processing that reduces refusal responses. The release is mainly intended for security research and safety testing.

Key points

1. 1

   The original model is a 426B (426 billion parameters) MoE (mixture of experts) architecture.

2. 2

   Quantization, a technique that stores numbers with less precision to reduce size, cut the model down to 112GB.

3. 3

   The full-precision version of the same model is also offered through a hosted API, and requests handled locally are not charged.

4. 4

   The 112GB figure is only the model itself, so extra memory needed to run it must be prepared separately. No independent performance re-measurement data is available yet.

Why it matters

As large models increasingly run on personal hardware, the fact that data does not leave external servers is drawing attention from organizations with strict security needs.