Meta opened the weights for Muse Glimmer on 10 August 2026 - a 30-billion-parameter dense model from the Meta Superintelligence Lab, built to run on consumer hardware. Mark Zuckerberg's announcement is short, and the model card is where the actual story sits: multimodal in (text and images), agentic by design, and quantised to fit a 24 GB card.
It's also one of several capable lightweight models out of the States right now, and the choice at this size is real. The Qwen 3.8-Max weight drop on 12 August 2026 is about to add a serious Chinese contender to the same shelf. More on that at the end.
Muse Glimmer 30B
The specs, plainly. Dense causal transformer, ~29.6B parameters total including a ~1.8B perception encoder (a ViT-G/14) for vision. 131,072-token context. Text and image in, text out. Apache 2.0. Knowledge cutoff 4 January 2026.
What makes it worth the download is the agentic framing. Meta trained and evaluated it on full-task agent benchmarks - MCP Atlas 75.5, SWE-Bench Verified 76.0, DeepSearch QA 74.6 - and the card lists scaffold compatibility with Hermes Agent and OpenClaw by name. That last bit matters if you're running an agent locally; a model tested against your scaffold is the safer bet for clean tool calls.
The local-deployment maths is the other half. Meta ships two 4-bit GGUF k-quants in the official GGUF repo:
muse-glimmer-30B-kquant-17gb.gguf(16.8 GB) - fits a 24 GB GPU with ~1.0% average degradation across 15 benchmarks.muse-glimmer-30B-kquant-dynamic.gguf(19.7 GB) - fits 32 GB with ~0.2% degradation.
The two text builds are text-only on their own. For image input, add the mmproj-kquant.gguf perception encoder (~1.4 GB). There's also an optional dflash-kquant.gguf speculative drafter (~1.6 GB) - Meta reports a 3.1x speedup on an RTX 5090 and 1.8x on an M5 Max when you use it. We've added both quants to the Recommended Models page.
Ornith 1.0
Ornith 1.0 from DeepReinforce AI is the frugal one. The 9B dense cut runs in 16 GB and carries a 262K context window - handy for a starter card or a latency-sensitive serving box. There's a 35B reinforcement-tuned MoE sibling if you've got 24 GB to spend. Neither is frontier, but both are honest workhorses for the footprint.
Liquid AI LFM2.5
Liquid AI is the architecture play. LFM2.5-8B-A1B uses a hybrid A1B design for high throughput at low cost - the kind of model that's cheap to keep warm and quick on the first token. 128K context, Q8_0 GGUF around 9 GB, runs comfortably in 16 GB. Liquid is US-based, if you're keeping track.
The bar Qwen sets
It's a good moment for open-weight AI, and the US models above are a genuine sign of where the field is going - runnable, capable, agentic. They're also not quite at the level the Chinese side has reached. Qwen3.6-27B, the existing 27B dense Qwen from April, already outscores Muse Glimmer on the agentic coding harnesses: TerminalBench 2.1 puts Qwen at 60.7 against Muse Glimmer's 51.7, and SWE-Bench Verified goes the same way (77.2 to 76.0). Those are Meta's own numbers. That's the bar today, and it's about to move again - Alibaba opens the Qwen 3.8-Max weights on 12 August 2026, 2.4T parameters on the Qwen 3.5 architecture, multimodal, Apache 2.0, with the Max-Preview already scoring well on Alibaba's Token Plan and on Qoder ahead of the drop.
We covered the broader July picture - DeepSeek V4-Flash, Kimi K3, and the rest - in our last roundup. The short version is that the choice on both sides of the Pacific has rarely been this good, and the lightweight end especially so.
A small comparison
| Model | Params | Context | Modality | VRAM tier | Licence | Weights |
|---|---|---|---|---|---|---|
| Muse Glimmer 30B (K-Quant-17GB) | 29.6B dense + 1.8B vision | 131K | Text + image | 24 GB | Apache 2.0 | HF |
| Ornith 1.0 9B | 9B dense | 262K | Text | 16 GB | MIT | HF |
| Liquid AI LFM2.5-8B-A1B | 8B hybrid | 128K | Text | 16 GB | LFM 1.0 | HF |
| Qwen 3.8 27B (expected) | 27B | TBD | TBD | TBD | TBD | TBD |
Specs from vendor releases. VRAM tier is our recommendation for the listed quant, not the model's absolute floor. The Qwen 3.8 27B row is speculative - Alibaba hasn't confirmed a 27B sibling to the 3.8-Max, so the specs are TBD pending any announcement.
The common thread: agents on modest hardware
The thing that ties these together isn't geography. It's that a 16 to 24 GB box now runs a properly capable agent model - multimodal in Muse's case, long-context in Ornith's, cheap-to-serve in Liquid's. If you're pointing Hermes Agent or a similar scaffold at your own hardware, that's the interesting part. Three of these models are explicitly listed or shaped for that workload, and they're a solid choice now even if Qwen 3.8-Max changes the picture when it lands.
Weights on disk still aren't an API
Same catch as ever. A model on disk isn't a service. To run these in production you want routing across machines, auth and keys, failover when a box disappears, and the OpenAI-shaped endpoint your tooling already speaks.
That's what Infersec is for. Run Infersec's Conduit application on your hardware - Macs with Apple Silicon, NVIDIA rigs, AMD accelerators - and it dials home, pulls the model config, and serves. Point an endpoint at your sources and out comes an OpenAI- and Anthropic-compatible API. HuggingFace integration fetches the weights, MCP handles tool calls server-side, and Infersec itself - operated by Vault Garden Oy, a Finnish company - processes customer data in the EEA on Scaleway infrastructure in the Netherlands. The Conduit application runs on your hardware, wherever that is. On the public inference path there's no prompt storage - prompts live in memory for the request and are gone when it ends.
Because the endpoint speaks OpenAI, most of what you'd point at a frontier model should work as-is - opencode, hermes, mastra, LangGraph. Swap the base URL for your Infersec endpoint, drop in a key, and you're on weights nobody can take back.
Sign up for early access, or browse the Recommended Models page.