In seven days, three open-weight frontier releases landed that together cover the whole local-AI stack. Meta’s Muse Spark Open came first on October 15, optimized for Apple silicon. Reflection’s Beam came three days later on October 16, sized for serious GPU hardware and distributed on Ollama with a temporary free cloud tier. IBM’s Granite Switch 4.1 is the smaller third entry, with 8B and 30B variants sized for a single consumer GPU. Together they map out the practical options a self-hosted reader faces right now.
Muse Spark Open: a 1-million-token model built around the Mac Studio
The headline numbers come from the Hugging Face release notes. It is a 62-layer mixture-of-experts with 256 experts per MoE layer, 8 routed per token. Total parameters sit at roughly 816 billion, with about 42 billion active per forward pass, and the context window stretches to a full 1,048,576 tokens (about 266 GB of KV cache at BF16). Modalities are text, image, audio, and video in; text out. BF16 weights land at about 1.5 TB on disk, NVFP4 at about 484 GB, and the community has produced GGUF (Q2_K, Q4_K_S, Q8) and MLX oQ (oQ3e, oQ4e) builds for local runners.
The Apple angle is concrete. TechCrunch reports that Meta “teamed up with Apple” for the launch and that the model “has been designed to work with Macs,” running best on a Mac Studio with the M5 Ultra chip. The Hugging Face team tested the oQ3e quant on an M3 Ultra with 512 GB of unified memory and recorded “about 225 tok/s prefill and 21 tok/s generation with a 1024-token prompt and 128 generated tokens,” with the model taking “about 344 GB” of RAM. The 2-bit imatrix-calibrated build, which the team ran at 18 tokens per second, used 304 GB. Local execution requires at least 1 TB of RAM, per TechCrunch.
The release notes describe a “Contemplating” reasoning mode with low / medium / high / xhigh effort levels mapped to 16 / 64 / 256 / 512 generated tokens, tool calling including multimodal tools, a Harmony channel chat template (self / commentary / user), and audio at 24 kHz mono at about 12.5 tokens per second. If you already own a maxed-out Mac Studio, Muse Spark Open is the first openly licensed frontier-tier multimodal you can run without leaving the local network.
Reflection Beam: a 501B MoE for Blackwell, free on Ollama for now
Per the Reflection launch blog and Hugging Face’s October 16 writeup, Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, 52 layers, 256 experts per MoE layer with 8 routed, a 1 million-token context window, and an Apache 2.0 license. The MoE holds 98.4% of the parameters, and the Hugging Face piece describes a depth-weighted residual stream that keeps the hidden state “at constant scale.”
The training run is large. Reflection says pretraining finished in under four weeks on 6,144 GB300 NVL72 GPUs with 92.3% goodput, with a separate reinforcement-learning deployment on 10.5K GB300 GPUs running for another four weeks. Their pipeline eliminated about 95% of raw internet tokens, kept roughly 1.8 trillion that “conventional filters would have missed,” and ran more than 100 million rollouts in roughly 1.3 billion sandboxes.
For local-AI readers, the practical question is hardware. The Hugging Face post gives concrete configs: BF16 needs about 1 TB on disk, FP8 is sized for Hopper and Ampere, NVFP4 brings BF16 down from 1 TB to about 280 GB. The post recommends NVFP4 on “4x B200 (720 GB)” or “2x B300,” and FP8 on “8x H200.” For everyone without Blackwell hardware, the post points to “Inference Providers/serverless.” Beam on its own hardware means a real rack.
That is what makes Ollama’s announcement the more interesting part. Ollama listed Beam as “freely available on Ollama for a limited time” via the beam:cloud tag, with launch hooks for Claude Code, Codex, and OpenCode. The blog describes Ollama Cloud as running Beam under “full zero data retention (ZDR).” Reflection’s launch blog describes Beam as competitive with larger open models like GLM-5.2 “while using 3-4x less inference compute.” If you do not have a GPU cluster, Ollama Cloud is currently the realistic way to run Beam without your prompts touching a closed lab.
Granite Switch 4.1: the one that runs on a single GPU
The third release is the one a self-hosted reader with a normal workstation can actually use. IBM Research’s Granite Switch 4.1 blog describes a modular “Switch” design built around Activated LoRAs (aLoRAs): a base Granite model plus adapters that include a Decision aLoRA, a Guardian aLoRA, and a Granite-Speech-5.0-Turbo-CTC ASR encoder. The blog ships two variants, “ibm-granite/granite-switch-4.1-8b” and “ibm-granite/granite-switch-4.1-30b,” and frames the smaller one as the practical default.
The notable claim is that the composed model takes “approximately the same amount of VRAM on a GPU as the base model would by itself.” The blog documents a 5-step RAG workflow completing faster than Jev-27B on the 30B variant, and a custom 3B Switch variant hitting “60 - 100 decisions / second” against “10 decisions / second” on Jev in a Doom demo. The 8B variant fits on a consumer card; the 30B variant needs a larger consumer card. For a single-GPU self-hosted box, Granite Switch 4.1 is the only one of these three that does not require a rethink of your hardware.
What This Means
A reader’s decision tree now has three real branches. If you own a Mac Studio with M5 Ultra and at least 1 TB of unified memory, Muse Spark Open is the only openly licensed frontier multimodal you can run locally today. If you own or rent a Blackwell GPU cluster, Beam is the open-weights frontier with Apache 2.0 attached. If you have a single consumer GPU, Granite Switch 4.1’s 8B is the only one of the three that fits without quantization drama, and the Switch adapter design lets you swap a Decision aLoRA in and out without swapping the weights.
The pricing shape matters too. Ollama’s temporary free tier with zero data retention, Beam-as-cloud, is a small but real change: it puts a 501B model behind the same privacy posture Ollama documents for its local server, which is a meaningful pressure point on closed frontier APIs that train or retain on chat inputs by default.
The bigger picture is that the open-weight release cadence has moved from “two or three frontier drops a year” to “two or three in a week.” None of this replaces Anthropic or OpenAI on absolute benchmark ceilings, but a self-hosted answer to a frontier question no longer requires a long wait.
The Bottom Line
Three open-weight frontier releases landed in seven days. Muse Spark Open runs on a top-spec Mac Studio; Reflection Beam needs a Blackwell rack but is free on Ollama Cloud for now under zero data retention; IBM Granite Switch 4.1’s 8B variant is the one that fits on a single consumer GPU. For a self-hosted reader, the answer to “what should I actually run” now depends less on which model is best and more on which hardware is already on your desk.