Mistral medium 3.5 128B, MLX 4bit, ~70 GB
On April 30th, someone in the LocalLLaMA community posted about running Mistral Medium 3.5 at 128 billion parameters, quantized to 4-bit precision through Apple's MLX framework, arriving at a footprint of roughly 70 gigabytes. That is the complete evidence base: one post, one data point, no official confirmation from Mistral AI, no independent benchmarking, no corroborating signals from anywhere else. The previous read on this story noted that the combination — model tier plus implied hardware — was what made the post worth a second look even without corroboration. A 70GB 4-bit MLX quantization means someone is running a frontier-class model on high-end consumer Apple Silicon, likely a Mac Studio or Mac Pro with 192GB unified memory. That is not a laboratory curiosity. That is a genuine workflow.
If confirmed, here is what this means. Mistral Medium 3.5 at 128B parameters represents serious model capability — not a toy, not a fine-tune of something smaller. Getting it to run locally via MLX at 4-bit quantization would mean that practitioners with high-end Apple Silicon hardware can run a genuinely competitive model entirely offline, with no API costs, no data leaving the machine, and inference speeds that MLX has proven competitive on for this class of hardware. The second-order effect matters more than the first: every time a model of this scale becomes locally runnable, the negotiating position of cloud API providers weakens a little. Enterprises with compliance constraints and developers who have been waiting for local capability to catch up to hosted performance both move closer to a viable alternative. Mistral's positioning — consistently releasing capable models with relatively permissive terms — makes this particularly pointed if it holds up.
Watch for an official MLX model card or Hugging Face listing from a verified Mistral source, and for independent inference benchmarks that would either validate or quietly bury what this single post is claiming.
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1sz1qer/mistralaimistralmedium35128b_hugging_face/"> <img alt="mistralai/Mistral-Medium-3.5-128B · Hugging
<table> <tr><td> <a href="https://www.reddit.com/r/singularity/comments/1szno73/mistral_medium_35_a_reliability_first_open_source/"> <img alt="Mistral Medium 3.5: A reliability
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1s1ewzj/mistral4small_uncensored_30gb_mac_only_mlx_studio/"> <img alt="Mistral-4-Small UNCENSORED - 30GB -
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1sycgzj/mistralmedium_35_128b_spotted/"> <img alt="Mistral-Medium 3.5 (128B) spotted ?"
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1t1a40e/mistralai_mistralmedium35128bq3_k_m_experiences/"> <img
<!-- SC_OFF --><div class="md"><p>All ggufs were broken, resulting in bad outputs, especially at long context.</p> <p>Anyway, it is fixed now: <a
<!-- SC_OFF --><div class="md"><p><a href="https://unsloth.ai/docs/models/mistral-3.5">https://unsloth.ai/docs/models/mistral-3.5</a></p> <p>"May 1, 2026 Update: We worked
<table> <tr><td> <a href="https://www.reddit.com/r/singularity/comments/1sz95za/mistral_medium_35_128b_is_launched/"> <img alt="Mistral Medium 3.5 128B is launched"
NewsHive monitors these sources continuously. All signal titles above link to the original reporting.
Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →