I got Mixtral-8x7B running on a M1 MBAir!!!
On March 23rd, someone arrived on LocalLLaMA with the specific excitement of a person who had watched a progress bar finish that had no business finishing. The claim: Mixtral-8x7B, Mistral AI's mixture-of-experts model carrying the weight of roughly 46.7 billion parameters, was running on an M1 MacBook Air. Not a workstation. Not a cloud VM with 80GB of HBM2. A fanless consumer laptop — one that shares a unified memory pool between CPU and GPU, throttles when it gets warm, and was not designed with anything like this in mind. The post scored 6.8 on LocalLLaMA, which is the community's way of saying: we're not convinced, but we're listening. That distinction matters. LocalLLaMA has seen enough extraordinary claims to develop a decent filter for them.
If confirmed, here is what this means. The practical barrier to running frontier-adjacent open models drops to something approaching zero for anyone who owns recent Apple silicon hardware. Mixtral-8x7B has been competitive with GPT-3.5 on several benchmarks — running it locally, privately, without API costs or rate limits, on hardware millions of people already own, is not a minor footnote. It changes the calculation for developers prototyping offline applications, for researchers in institutions with restricted cloud budgets, for anyone whose use case involves data that cannot leave the device. The second-order effect is pressure: if a fanless laptop can run this, the justification for cloud dependency on models of this size gets harder to defend. Apple's unified memory architecture, long treated as a curiosity in ML circles, gets a louder argument made on its behalf.
Watch for a follow-up post with actual inference speeds, memory usage figures, and the quantization method used — those numbers would either make this a legitimate milestone or explain away the headline entirely.
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1srhtvg/where_we_are_in_a_year_everything_has_changed/"> <img alt="Where we are. In a year, everything has
<!-- SC_OFF --><div class="md"><p>One thing we’ve been noticing lately is that a surprisingly large percentage of day-to-day AI workflows no longer seem to require frontier-scale
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1s7xv6u/got_a_9b_abliterated_claudedistilled_model/"> <img alt="Got a 9B Abliterated Claude-Distilled model
<!-- SC_OFF --><div class="md"><blockquote> <p><strong>WARNING: wait before download from HF: I just realised my upload of the new versions with the additional fix in the chat
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1s59rct/rys_part_3_llms_think_in_geometry_not_language/"> <img alt="RYS Part 3: LLMs think in geometry, not
<!-- SC_OFF --><div class="md"><p>Just putting together a quick list of the new <strong>open-source physical AI / robotics models</strong> from <strong>NVIDIA GTC
<!-- SC_OFF --><div class="md"><p>Link to the paper:</p> <p><a
<!-- SC_OFF --><div class="md"><p>I have experience using llama.cpp on Windows/Linux with 8GB NVIDIA card (384 GB/s bandwidth) and offloading to CPU to run MoE models. I typically
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 →