--- title: Meta releases Muse Glimmer, a 30B open-weights model... description: Meta has released Muse Glimmer, a 30B parameter dense model under the Apache 2.0 license, making it freely usable, modifiable, and distributable with minimal... canonical: https://daily.dev/posts/meta-releases-muse-glimmer-a-30b-open-weights-model-under-apache-2-0-scxtb9qhg twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Meta releases Muse Glimmer, a 30B open-weights model under Apache 2.0 | daily.dev og:description: Meta has released Muse Glimmer, a 30B parameter dense model under the Apache 2.0 license, making it freely usable, modifiable, and distributable with minimal... og:url: https://daily.dev/posts/meta-releases-muse-glimmer-a-30b-open-weights-model-under-apache-2-0-scxtb9qhg og:image: https://api.daily.dev/og/posts/sCxtB9qHg.png og:image:alt: Meta releases Muse Glimmer, a 30B open-weights model under Apache 2.0 og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Meta releases Muse Glimmer, a 30B open-weights model under Apache 2.0 **[Collections](https://daily.dev/sources/collections)** · 7 min read · 1 upvotes · 0 comments ## Summary Meta has released Muse Glimmer, a 30B parameter dense model under the Apache 2.0 license, making it freely usable, modifiable, and distributable with minimal restrictions. The release comes from Meta's MSL team, led by Scale AI's Alexandr Wang. Mark Zuckerberg also announced that weights for Muse Spark 1.2, described as Meta's latest foundation model, will be released soon. The Apache 2.0 licensing is highlighted as a meaningful distinction from more restrictive open-weight releases. ## Content Meta just did something it hasn't done in a while: shipped an open-weight model people actually seem excited about. Muse Glimmer is a 30-billion-parameter dense model, licensed Apache 2.0, built specifically to run agentic workloads on a single consumer GPU or a beefy laptop. No cloud API, no metering, no waiting on rate limits. Given how rocky Llama 4's reception was and how much drama has surrounded Meta's AI reorg since, this feels like a genuine reset. ## What it actually is Glimmer is distilled from Meta Superintelligence Labs' larger, closed model, Muse Spark, using a pipeline that combines logit distillation, supervised fine-tuning, reinforcement learning, and on-policy distillation. Meta explicitly places Glimmer outside its own "Frontier AI" definition, since it's meaningfully weaker than Spark. Spark itself is only reachable through Meta's API — Zuckerberg has said an open version, Muse Spark 1.2, is coming "soon," reportedly benchmarking close to Claude Opus 4.8. Architecturally, Glimmer looks a lot like a tuned-up Gemma. Andrej Karpathy-style close readers (well, Sebastian Raschka in this case) pointed out the family resemblance immediately: Gemma-style pre/post RMSNorm placement, but with SwiGLU instead of GeGLU, gated attention on both grouped-query and sliding-window attention, and a genuinely extreme GQA ratio — 32 query heads to just 2 KV heads. The payoff is a tiny KV cache: 52 KiB per token in BF16, versus 64 KiB for Qwen3.6 27B and a whopping 840 KiB for Gemma 4 31B. Context tops out at 131K+ tokens, which is shorter than Qwen3.6 or Gemma 4's native windows — a reasonable but noticeable tradeoff for an agent-focused model in 2025. ## Fitting a 30B model on your desk At full precision this thing needs 55GB+ of memory, which rules out most consumer setups. Meta's answer is 4-bit quantization, which shrinks it to under 20GB — small enough to leave room in a 24–32GB GPU for the KV cache, a perception encoder for screenshots, and a speculative decoding drafter, all at once. There's a K-Quant-17GB build for 24GB systems and a heavier dynamic-quant variant for 32GB setups. People have gotten it running on an RTX 3090/4090/5090, AMD 9700, and a MacBook Pro with 64GB RAM. Unsloth even got a 2-bit GGUF calling 100+ tools inside 14GB of RAM, which is the kind of thing that sounded implausible a year ago. Speed comes largely from DFlash, a small draft model that guesses a full block of tokens ahead while the main model verifies the block in one pass instead of generating token by token. On an RTX 5090 this pushed decode speed from 74.9 to 233 tokens/sec — a 3.1x jump, measured through llama.cpp. Apple Silicon sees smaller but still solid gains: 1.8x on an M5 Max, 1.5x on an M4 Max, landing around 50 and 38 tokens/sec respectively. ExecuTorch's writeup adds more color on the backend work involved — CUDA graphs, packed K-quant kernels, and FlashDecoding++ on the NVIDIA side; RMSNorm, RoPE, and SDPA lowered to native Metal ops on Mac. There's also a hybrid global/sliding-window KV cache where only 13 of 52 layers are global, which is presumably part of why the memory footprint stays so low.// ## How it stacks up On Meta's own numbers, Glimmer leads Gemma4-31B and Qwen3.6-27B on MCP Atlas (75.5 vs 54.2 and 62.5). That lead flips on computer-use and terminal coding benchmarks — Qwen3.6-27B comes out ahead on OSWorld-Verified, TerminalBench 2.1, and SWE-Bench Verified. Independent composite scores (Artificial Analysis Intelligence Index) put Glimmer slightly behind Qwen3.6 overall, so it's genuinely close, not a blowout in either direction.// Real-world testing backs up the mixed picture: in one arcade-game test on RTX 5090 hardware, Glimmer's Space Invaders, Tetris, and Arkanoid implementations played correctly where Gemma 4 31B's didn't — but it burned 83.4K tokens and 17.7 minutes to Gemma's 17.8K tokens and 4.5 minutes. That's the tradeoff in a nutshell: Glimmer tends to think longer and use more tool calls, which sometimes pays off and sometimes just costs you tokens. A separate comparison against its own teacher, Muse Spark 1.2, found Glimmer generating 5 runnable games at roughly a fifth of Spark's API cost ($0.02 vs $0.11), though Spark's graphics looked better and Glimmer stumbled more on the racing game specifically. In agentic coding tests against Qwen3 27B run through a local coding agent, Glimmer made more tool calls (57 vs 50) and produced more feature suggestions, using 36% more input tokens and 5.6% more output tokens along the way — plausibly a result of training that leaned hard into long-horizon execution and failure recovery. One fine-tuning note worth flagging: Hugging Face's Merve found that despite topping ScreenSpot-Pro, Glimmer's zero-shot performance struggles with the kind of ambiguous, human-style prompts in AI2's MolmoWeb dataset (things like "jump to nutrition facts"). Base model click accuracy was 13%, with 35% landing within 5% of the target. After fine-tuning on MolmoWeb data, that jumped to 41% accuracy and 68% within 5%. So it's a strong base, but not magic out of the box for every UI-grounding task. ## Security is a real consideration, not an afterthought Meta reports a 28.4% attack success rate on Siren AgentDojo and a 26.4% violation rate on CI Memories for Glimmer — numbers Meta itself flags as reasons to add safeguards rather than deploy it as a standalone endpoint. Other reporting cites a 28% prompt injection success rate for Glimmer versus 40% for Qwen, which — if accurate — is a mild win, but 28% is still not a number I'd want to bet a production agent's safety on unsandboxed. ## Where you can actually run it Support landed fast and wide. Day-zero integrations include llama.cpp, MLX, ExecuTorch, Ollama (0.32.7 shipped with it built in), LM Studio, vLLM, and SGLang, plus Hugging Face transformers. Red Hat AI released an FP8-block quantized checkpoint at roughly half the memory and disk footprint. Intel's LLM-Scaler added same-day support with FP8 online quantization on the Arc Pro B70, and AMD's Lemonade SDK 11.6 folded it into its model catalog with draft decoding, vision, and tool calling via llama.cpp. NVIDIA's own writeup highlights over 20K tokens/sec on a single GPU across RTX 5090, DGX Spark, DGX Station, and Jetson edge hardware, with fine-tuning support through NeMo AutoModel (SFT, LoRA) and NeMo RL for reinforcement learning. Unsloth has a free fine-tuning notebook that supports GRPO RL training too. ## The bigger picture, and my honest read Zuckerberg tied the release to a 6,500-word essay, "The Future Is for Everyone," arguing against concentrating superintelligence in a handful of labs. A couple of lines stood out to me: "Any policy that slows American model releases — even by a month — could add significant risk to American leadership," and a warning that the "most dangerous scenario" might be labs training powerful models and simply keeping them for themselves. He also announced Muse Spark 1.2's weights are coming, and hinted at even bigger models down the line — possibly connected to whatever internally is being called "Watermelon." I find myself genuinely torn on how to read all this. On one hand, an Apache 2.0 licensed 30B model that runs well on a single GPU, with real day-zero tooling support across the entire ecosystem, is unambiguously good for anyone who wants to build local agents without cloud dependency. That's not nothing. On the other hand, it's hard to ignore that the actually powerful model — Muse Spark — stays locked behind an API, and Glimmer represents the ceiling of what Meta is currently willing to let people own outright. Whether that's a meaningful step toward decentralized AI or a well-timed PR move alongside a company facing plenty of separate controversies is a matter of where you already stood on Meta before this happened. Probably both things are true at once. What I keep coming back to, though, is just how good the engineering is regardless of the politics. A 30B dense model with a 52 KiB per-token KV cache, real multimodal support, coherent long-context behavior, and speculative decoding tight enough to triple throughput on consumer GPU — that's a genuinely useful artifact to have in the open, whatever Meta's motives for releasing it. ## Questions this post answers ### What license is Meta's Muse Glimmer 30B model released under? Muse Glimmer is released under the Apache 2.0 license, which allows developers to use, modify, and distribute the model with relatively few restrictions. This is a meaningful distinction from many open-weight releases that impose usage conditions, making it one of the more permissive licenses available for a model of this scale. _Developers evaluating open-weight models for commercial use track licensing details like these on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 113 comments across hackernews (as of 2026-08-15). **TL;DR:** Developers are cautiously optimistic about Muse Glimmer's open-weights release and competitive benchmarks, but raise real concerns about hardware requirements, inference speed on consumer GPUs, and whether it can hold its own against the imminent Qwen3.8 27B. **Sentiment:** 45% positive · 35% mixed · 20% skeptical **The case for** - Apache 2.0 licensing is more permissive than previous Llama licenses, setting a positive precedent for the industry. - Pre-quantized 4-bit weights (~20 GB) and a bundled DFlash speculative-decoding drafter make local deployment more practical out of the box. - Benchmark comparisons against Gemma 4 and Qwen3.6 look competitive, especially on tool-calling tasks. - Privacy and cost-control benefits of running a capable model fully locally are meaningful for many developers. - Multi-token prediction (MTP) helps dense models approach MoE speeds without sacrificing intelligence. **The pushback** - Effective memory footprint including KV cache, perception encoder, and drafter likely exhausts a 24 GB GPU at long context, which is exactly the agentic workload it targets. - Dense architecture means slow decode on DDR5/unified-memory systems (~15 tok/s on a DGX Spark), unlike MoE alternatives. - Benchmarks compare against Qwen3.6 27B (4 months old), and Qwen3.8 27B is expected to surpass it imminently. - Hardware capable of running it comfortably (64 GB Mac or 24 GB GPU laptop) costs $3,500–$4,000+, limiting accessibility. - "Minimal to no degradation" from 4-bit quantization on tool-call schema adherence is an unverified strong claim with no per-quant breakdown provided. - Energy costs of local inference likely mean users never break even versus cheap API providers. **By community** - hackernews (mixed): Broadly welcoming of the open-weights release and Apache 2.0 license, but skeptical about real-world hardware requirements, inference speed, benchmark honesty, and whether Qwen3.8 will immediately obsolete it. **Hottest debate:** Whether the hardware cost and inference speed of a dense 30B model make local deployment practical compared to cheap cloud API alternatives or MoE models of similar size. **Open questions** - How does Muse Glimmer perform at 64k+ context lengths on a 24 GB GPU once KV cache, encoder, and drafter are all resident? - Will Qwen3.8 27B surpass Glimmer-30B on most benchmarks when it releases? - Does the 4-bit quantization cause measurable degradation in tool-call schema adherence (malformed JSON, wrong enum values) that benchmark averages don't capture? - Is the open-weight version of Muse Spark 1.2 identical to what Meta serves internally? **Highlights** > The memory math is the part I keep re-reading. 4-bit gets the LM under 20GB, but they're explicitly budgeting the KV cache, the perception encoder, and the DFlash drafter into the same 24GB envelope. On a 4090 that leaves maybe 3-4GB of KV once the drafter and encoder are resident, so long agent traces are going to spill or truncate — which is exactly the workload this model is trained for. I'd like to see the K-Quant-17GB numbers reported at 64k+ context, not just conversational length. "Minimal to no degradation on agentic tasks" from quantization is also a strong claim. In my experience 4-bit shows up first in tool-call schema adherence — malformed JSON args, wrong enum values — before it moves benchmark averages. Does the report break down tau-Bench / MCP-Atlas per quant level? > — [jkwang on hackernews](https://news.ycombinator.com/item?id=49242100) > Wow, Meta is back (at least for now)! I like this class of model. Multi-token prediction makes it viable to run dense models at not-too-far-off speeds as MoE models with much better intelligence. The submission’s title (open weights 30B local coding model) is luckily wrong: This is meant to be a general agentic model. It even comes pre-quantized and with a MTP/drafter model. Looking good! Let’s hope they aren’t dishonest with the benchmarks this time … > — [solarkraft on hackernews](https://news.ycombinator.com/item?id=49242071) > It is interesting but it does look like a careful distillation of (Spark and) biggers open-weight models. The progress compared to Qwen3.6 27B is good, not that impressive, it's a 4 months old model. (kuto to them to compare to 27B dense and not 35B MoE, it's more fair to do so). It is very probable that Qwen3.8 27B will crush Glimmer-30B on most benchmarks. > — [\_ache\_ on hackernews](https://news.ycombinator.com/item?id=49242292) > This model I think will be too slow for that on Spark, even at 4 bit quant. It's a dense model, not MoE like e.g. Qwen 35b. On a Spark it will be memory bandwidth limited (unlike Gemma 4.) I haven't tried yet (working on it) but back of the napkin estimate puts it at around 15tok/s even after converting to NVFP4. Prefill would be much higher though. That 15tok/sec is pretty typical for dense models of this size. NVFP4 Q/K/V/O and MLP projections: ~13 GB/token BF16 attention gates: ~3 GB/token BF16 LM head: ~2.5 GB/token Total: ~18.9 GB/token At 273 GB/s, that gives a bandwidth-only ceiling of about 14.5 tok/s; actual performance would be lower. > — [cmrdporcupine on hackernews · 1 comments](https://news.ycombinator.com/item?id=49242179) > Some interesting findings from the chat template designs: 1. The template name is Onyx ATEM as found in the tool call exception message 2. It appears to be following a harmony-style chat template. But the tool use seems to be a xml like : / / 3. atem: a internal joke of meta in reverse? https://huggingface.co/meta-models/Muse-Glimmer-30B/blob/mai... > — [polymorph1sm on hackernews · 1 comments](https://news.ycombinator.com/item?id=49242626) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49241679) · 268 points · 113 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#llm](https://daily.dev/tags/llm) [View this post on daily.dev](https://daily.dev/posts/meta-releases-muse-glimmer-a-30b-open-weights-model-under-apache-2-0-scxtb9qhg) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Meta releases Muse Glimmer, a 30B open-weights model under Apache 2.0","url":"https://daily.dev/posts/meta-releases-muse-glimmer-a-30b-open-weights-model-under-apache-2-0-scxtb9qhg","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/meta-releases-muse-glimmer-a-30b-open-weights-model-under-apache-2-0-scxtb9qhg"},"datePublished":"2026-08-10T10:10:58.924Z","dateModified":"2026-08-15T09:44:06.444Z","description":"Meta has released Muse Glimmer, a 30B parameter dense model under the Apache 2.0 license, making it freely usable, modifiable, and distributable with minimal...","image":"https://pbs.twimg.com/media/HPWmJnGakAAp5zn.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HPWmJnGakAAp5zn.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/meta-releases-muse-glimmer-a-30b-open-weights-model-under-apache-2-0-scxtb9qhg","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,llm","timeRequired":"PT7M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Meta releases Muse Glimmer, a 30B open-weights model under Apache 2.0"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/meta-releases-muse-glimmer-a-30b-open-weights-model-under-apache-2-0-scxtb9qhg#faq","mainEntity":[{"@type":"Question","name":"What license is Meta's Muse Glimmer 30B model released under?","acceptedAnswer":{"@type":"Answer","text":"Muse Glimmer is released under the Apache 2.0 license, which allows developers to use, modify, and distribute the model with relatively few restrictions. This is a meaningful distinction from many open-weight releases that impose usage conditions, making it one of the more permissive licenses available for a model of this scale. Developers evaluating open-weight models for commercial use track licensing details like these on daily.dev."}}]} ```