# Muse Glimmer 30B on a Mid-Range Home Build > Source: > Published: 2026-08-12 18:01:01+00:00 *Meta’s first Superintelligence Labs model shipped this morning. By the afternoon it was generating tokens on a six-year-old Ryzen with two consumer GPUs. Here is what it actually does — and the three walls you hit getting there.* Meta Superintelligence Labs [released Muse Glimmer](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model): 30 billion parameters, dense rather than mixture-of-experts, Apache 2.0, and built specifically for local agents — multi-step reasoning, tool calls with strict schemas, image understanding, and failure recovery. It is distilled from Muse Spark, its much larger teacher, through logit distillation followed by agent-heavy mid-training and RL post-training. It ships with a frozen ViT-G/14 perception encoder and a separate speculative drafter based on [DFlash](https://arxiv.org/abs/2602.06036), and the weights are on [Hugging Face](https://huggingface.co/meta-models/Muse-Glimmer-30B). The official performance numbers were measured on an RTX 5090. Most people don’t have one. This is what the model does on hardware people actually own. Nothing here is exotic. The two cards were bought separately, a generation apart, and neither was ever a flagship. Meta’s own guidance is a 24–32 GB memory envelope: over 55 GB at full precision, under 20 GB once quantised to roughly 4-bit — their official variant is called K-Quant-17GB — leaving headroom for the KV cache, the perception encoder and the drafter. This machine sits squarely inside that window, which makes it a fair test of the claim rather than a stress test of it. **01 — Ollama will not run it.** Ollama shipped a release the same morning naming Muse Glimmer explicitly. It still rejects every GGUF variant: only the Apple-silicon MLX build is accepted, because NVIDIA support hasn’t landed yet. If you’re on a PC, the packaged route is a dead end. Build llama.cpp from source is necessary. **02 — CUDA 13.1 no longer compiles on Ubuntu 26.04.** The build fails before it starts: the updated glibc declares rsqrt with an exception specification incompatible with NVIDIA's own math headers. This is not a compiler-version problem — the misleading "gcc versions later than 15 are not supported" message sends you chasing the wrong thing, and forcing gcc 14 produces exactly the same failure. The fix is to build inside a container on an older glibc. The resulting binaries run natively on the host, since glibc is forward-compatible. **03 — Without ****--jinja the model never answers.** Left to llama.cpp's built-in chat template, Muse Glimmer echoes the prompt and never closes its reasoning block. Every token lands in reasoning_content and the content field stays empty — it looks like a broken model rather than a template mismatch. --jinja makes the server use the template embedded in the GGUF, and the behaviour becomes correct immediately. The RTX 5090 and Apple figures are Meta’s published numbers. Everything in the right-hand column was measured locally with llama-bench and the server’s own timings output. One caveat before comparing across columns: Meta’s chart footnote says the Apple figures were measured with ExecuTorch, while the 5090 — like this machine — ran llama.cpp. The official runs also used the K-Quant-17GB model with the quantised drafter, where this test used a community 4-bit GGUF. A speedup multiplier travels across runtimes and quantisation schemes better than an absolute tok/s figure does, but it is not a clean comparison, and part of the gap between 3.1× and the Apple band could be implementation rather than silicon. The llama.cpp-to-llama.cpp comparison — 5090 versus this pair — is the one that carries the argument below. A 3.1× gain on a 5090 and a 1.6× gain on two mid-range cards is not a contradiction, and it isn’t a sign that something is misconfigured. It follows from what speculative decoding actually does. The drafter proposes a block of tokens; the large model then verifies the whole block in a single pass. That verification batches work that would otherwise have been a sequence of separate, memory-bound steps — which shifts the bottleneck away from memory bandwidth and toward raw compute. Hardware with compute headroom to spare converts more of the proposal into real speedup. An RTX 5090 has roughly 3.5× the memory bandwidth of this pair and several times the compute, so it has far more headroom to spend. Draft acceptance here sat between 50 and 62 percent, which is healthy. The proposals were good; the machine simply couldn’t verify them as cheaply. There is a second thing worth reading off Meta’s own chart. The 233 tok/s is a mean over seven prompt categories, and the range bar around it is enormous — roughly 130 to 335 tok/s depending on the category. The baseline bars carry almost no spread, which is exactly what you’d expect: plain decode doesn’t care what kind of prompt it is, but draft acceptance does. So even on the flagship, the 3.1× is an average hiding prompt types where the gain is far smaller. The 40–47 tok/s range measured here is the same phenomenon at a smaller scale. Meta also published two figures that make the point better than the 5090 does: 1.5× on an M4 Max and 1.8× on an M5 Max — and in absolute terms, 38 and 50 tok/s with the drafter. Two mid-range desktop GPUs land inside that band, not near the flagship (and, for what it’s worth, ahead of an M5 Max on raw decode). The 3.1× is the outlier in Meta’s own table — it just happens to be the number that made the headline. A 1.5× to 1.7× speedup is the honest number for mid-range hardware. It is a solid gain — but if you budget for a tripling because that’s what the launch post said, you will be disappointed by physics, not by the model. Not generation — prompt processing at 1113 tok/s. On the same machine, an 80-billion-parameter mixture-of-experts model processes prompts at 256 tok/s. Muse Glimmer, with barely a third of the parameters, is more than four times faster at reading. That gap has nothing to do with one architecture being cleverer than the other. It is a memory threshold. At 20.3 GB, Muse Glimmer fits entirely inside 28 GB of VRAM. The MoE does not: its routed experts are offloaded and computed on the CPU, against system RAM measured at 22.8 GB/s — against roughly 500 GB/s of VRAM bandwidth on the cards. Twenty-two versus five hundred. That ratio, not the parameter count, is the whole story of local inference. A model that fits is in a different performance class from a model that almost fits, and the boundary is sharp rather than gradual. Performance articles rarely check whether the thing is any good. Given a synthetic test image — three coloured shapes and two lines of small text — the model returned the shapes with correct colours and positions and transcribed both lines exactly. The perception encoder costs about 1.9 GB on top of the weights. The GGUF repository was created at 09:11 UTC with zero downloads. The runtime that can execute it needed compiling from a commit merged hours earlier. By late afternoon the model was answering questions, reading images, and running speculative decoding on a desktop assembled from parts. The rough edges are real — a packaged runtime that refuses the file, a toolchain that no longer builds on a current distribution, a flag without which the model appears broken. None of this contradicts the launch post, which says optimized integrations for llama.cpp, MLX and ExecuTorch “will land in the coming days”. Day one is simply earlier than the coming days. All three walls are the cost of being early, and all three will be gone in a fortnight. What won’t change is the more interesting finding. Meta’s headline speedup does not transfer to ordinary hardware — and even in Meta’s own chart it is an average over prompt categories with a spread wide enough to swallow the claim. The reason it doesn’t transfer is structural rather than incidental. If you run models locally, the number worth internalising isn’t 3.1× — it’s the distance between 22.8 and 500. *A note on method: the ideas, the test setup, and every measurement in this article are mine. AI (Claude by Anthropic) was used as a collaborator for structuring arguments and drafting prose. Every claim and every number has been reviewed, edited, and validated by the author. The irony of using AI to write about running AI on your own desk is not lost on me.* [Muse Glimmer 30B on a Mid-Range Home Build](https://pub.towardsai.net/muse-glimmer-30b-on-a-mid-range-home-build-2fee8fc9730c) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.