What happened
Meta published Muse Glimmer in August 2026 as an Apache 2.0-licensed, open-weight multimodal model designed to run AI-agent workflows on a user's own computer.
Meta's model card describes Muse Glimmer as a dense causal transformer with about 29.6 billion parameters, including a roughly 1.8-billion-parameter perception encoder. It accepts interleaved text and images and produces text, with a stated context length of at least 131,072 tokens. Meta says the model was distilled from its larger Muse Spark system and trained for multi-step planning, schema-based tool calls, coding, failure recovery, and use inside agent frameworks such as OpenClaw and Hermes Agent. The release does not add audio output, and video is handled as separate image frames rather than as a video-native modality.
The released artifacts include full-precision BF16 weights, two roughly four-bit quantized versions, the vision encoder, and a small DFlash drafter model for speculative decoding. Meta says the smaller quantized package compresses the language model below 20 GB so the model, its working-memory cache, its perception encoder, and the drafter can fit within a 24 GB or 32 GB memory envelope. Full precision targets 64 GB of accelerator memory. Those figures describe Meta's tested configurations, not a guarantee that every consumer laptop or graphics card will load the same context length or achieve usable speed.
Meta reports that DFlash proposes blocks of 16 tokens, which the main model verifies in parallel instead of generating every token one at a time. In company tests using greedy decoding and a batch size of one, the 17 GB quantized model rose from 74.9 to 233.4 tokens per second on an Nvidia RTX 5090, from 23.7 to 37.8 on an Apple M4 Max, and from 26.6 to 50.2 on an Apple M5 Max. Meta attributes average benchmark degradation of 0.2% to its dynamic quantization and 1.0% to the 17 GB version across 15 tests; independent laboratories have not yet reproduced those aggregate claims.
The public model repository is not gated and carries the Apache 2.0 license for commercial and research use. Its card says training drew on publicly available material, third-party data, information from Meta products and services, and material curated or enriched by vendors and Meta staff. It gives a January 4, 2026 knowledge cutoff and says more than 100 languages appeared in training, but it does not publish a complete dataset inventory or performance results for every language. The available download is specifically Muse Glimmer: references to the larger Muse Spark system describe the source used for distillation and comparative safety reasoning, not an additional set of Spark weights in this release.
Read the primary source: Meta's Muse Glimmer model card and released weights ↗
Why it matters
Muse Glimmer tests whether useful multimodal agents can move from cloud-only services to hardware that individuals and small organizations can control directly.
Local execution can change cost, privacy, and continuity. A developer can keep source code, images, and documents on a machine rather than sending each step to a hosted model, and an agent can continue working when a network service is unavailable. Open weights also allow researchers to inspect, adapt, and evaluate the system without relying on a provider's application interface. Those advantages are conditional: an agent still needs software that securely exposes files and tools, sufficient memory for the chosen context, and people who understand what information the surrounding application logs or transmits.
Meta's own benchmark table suggests a specialized profile rather than across-the-board leadership. The company reports 75.5 on MCP Atlas, 74.6 on DeepSearch QA, and 51.2 on SWE-Bench Pro, ahead of the Gemma4-31B and Qwen3.6-27B comparison configurations it used. On other tests, Qwen's reported results were higher: 75.6 versus Glimmer's 65.9 on OSWorld-Verified, 60.7 versus 51.7 on TerminalBench 2.1, and 77.2 versus 76.0 on SWE-Bench Verified. The comparisons use Meta's selected scaffolds, reasoning modes, and evaluation recipes, so they show where the company says Glimmer is competitive, not how every deployment will behave.
The license and device target lower an important practical barrier for nonprofits, universities, independent developers, and smaller firms that cannot operate a frontier-scale cloud model. They could fine-tune the released weights, test an agent against their own documents, or run repeatable experiments on fixed software. Yet a 24 GB accelerator is still expensive, long contexts consume additional memory, and the model card's tested throughput uses recent high-end hardware. Calling the model 'local' therefore does not mean it will run well on an ordinary office laptop, phone, or older graphics card.
Local control also moves responsibility toward the deployer. Meta recommends human confirmation before irreversible actions, data minimization, scaffold boundaries, and application-specific guardrails. That matters because an agent with permission to edit files, send messages, or operate a browser can cause harm through an inaccurate plan even if the underlying weights never leave the machine. The open release makes scrutiny and customization easier, but it also lets users remove surrounding protections. The public benefit will depend on whether agent frameworks make limited permissions and visible action histories the default rather than optional expert settings.
What to watch next
Watch for independent hardware tests, reproducible agent benchmarks, direct safety evaluation, and versioned disclosures as Meta updates the released model and its documentation.
The first test is reproducibility. Independent evaluators should publish exact prompts, tool schemas, scaffold versions, reasoning budgets, context lengths, and failure criteria for the agent benchmarks in Meta's table. They should also measure completed tasks, not just well-formed tool calls, and report the cost of retries, hallucinated actions, and human interventions. Hardware reviewers need to separate model-file size from total memory use, test the released quantized artifacts on 24 GB and 32 GB devices, and verify whether DFlash speed gains persist under sampling, long documents, images, and multi-tool workloads rather than short greedy-decoding prompts.
The safety evidence is notably more cautious than the product positioning. Meta assigns Glimmer 'Moderate or lower' chemical and biological risk, while its cyber and loss-of-control ratings are inferred from Glimmer being broadly weaker than Muse Spark 1.0 rather than established by a complete new public evaluation. On two agent-security measures, the card reports a 28.4% prompt-injection attack success rate on Siren AgentDojo and a 26.4 contextual-information-flow violation rate on CI Memories. Those figures depend on the test setup, but they are a direct warning that local operation does not eliminate unsafe instructions or inappropriate information sharing.
Developers should also test the limitations Meta lists: inaccurate or biased answers, weaker behavior on languages that were not deeply evaluated, errors in unfamiliar multi-step situations, edge-case changes caused by quantization, and no native audio or video understanding. The card says the model is not intended for download or use by people under 18 and places responsibility on deployers when minors may encounter it. Useful follow-up evidence would include multilingual audits, accessibility testing, bias analysis for image inputs, and clear disclosure of how data from Meta products and external vendors contributed to training.
Finally, the release needs a durable record of change. The repository can update weights, configuration files, usage policies, quantization recipes, and benchmark documentation after publication, while downstream packages may preserve older snapshots. Meta should attach versioned evaluation results and safety notes to substantive revisions so deployers can tell whether a change affects model behavior or only packaging. Researchers should pin exact repository revisions, compare updates on matched tasks, and document whether later versions produce measurable user value, larger hardware requirements, or different rates of prompt injection and irreversible mistakes.
