When running large models on-device, memory becomes the obstacle before performance does. On August 10, Meta released "Muse Glimmer," distilled using outputs from Muse Spark. By 4-bit quantizing an approximately 29.6-billion-parameter model, Meta demonstrated a design that runs the model itself along with its surrounding components within a 24GB or 32GB class memory budget. This release combines weight-lightening techniques with DFlash, which speeds up generation.
Muse Glimmer is designed for agents that handle on-device schedules and messages, as well as file and screen context. Meta explains that local execution allows the model to run without a network connection. This differs from designs that offload processing to the cloud in terms of where data is handled.
Fitting a 30B-class model into a 24GB/32GB memory budget
Muse Glimmer combines a dense causal transformer with a dedicated perception encoder. The total parameter count is approximately 29.6B, of which the vision encoder accounts for approximately 1.8B. It takes text and images as input and outputs text. The context length is 131,072 or greater, and the knowledge cutoff is January 4, 2026.
Meta explains that the language model portion, which would exceed 55GB at full precision, was quantized to approximately 4-bit, bringing it under 20GB. The released assets include full BF16 weights, as well as 4-bit quantized weights for 24GB and 32GB targets, a DFlash drafter head, and a frozen ViT-G/14 perception encoder. All are distributed under Apache 2.0, but what's released here is open weights—this does not mean the training data has also been made public.
However, the 24GB or 32GB figures are not simply the memory required for weights alone. Long inputs increase the KV cache, and handling images also engages the perception encoder. The DFlash drafter also consumes memory concurrently. Therefore, the claim that the model runs on a single GPU is premised on this specific memory budget and configuration. It does not mean there is headroom on any GPU or at any context length.
DFlash speeds up generation but doesn't shrink the weights
DFlash is not a technique for reducing the main model's memory usage. A drafter model proposes blocks of 16 tokens at once, and the main model verifies these candidates in parallel. If a proposal is accepted, the output advances further than if the main model generated tokens one at a time.
In Meta's model card, measurements combining K-Quant-17GB with quantized DFlash showed the RTX 5090 improving from 74.9 tok/s to 233.4 tok/s. The M4 Max went from 23.7 tok/s to 37.8 tok/s, and the M5 Max went from 26.6 tok/s to 50.2 tok/s. Meta reports these as 3.1x, 1.5x, and 1.8x improvements respectively.
These figures are average values measured using llama.cpp on the RTX 5090, and ExecuTorch on the M4 Max and M5 Max, with batch size 1 and greedy decoding. Speed will vary with different input lengths or sampling conditions. Quantization lowers the memory barrier, while DFlash shortens generation latency—these two roles need to be read separately.
Distilling Muse Spark's outputs to the device side
For training Muse Glimmer, Meta performed logit distillation using the logits output by Muse Spark as a teacher signal. In the subsequent mid-training phase, training focused on long-form text and agentic use, and post-training combined SFT, on-policy distillation, and reinforcement learning. This process trains a model that can run on-device from the output distribution of a teacher model.
Use cases span local AI agents, coding agents, and function calling. The model can also read images, screenshots, and documents. Meta explains it was trained on data in over 100 languages. Meta describes compatibility with agent configurations like OpenClaw and tool calling, and the model has been trained to retry when tool calls fail. Reasoning intensity can be selected from low to xhigh. It does not support voice input/output, and video is handled as individual frames.
Moving processing to the device side does not automatically bring agent safety along with it. The model card lists errors in multi-step reasoning, differences in edge cases due to quantization, and degraded performance in languages that haven't been sufficiently evaluated as limitations. For irreversible operations, human confirmation is required, and countermeasures against things like prompt injection are still needed at the system level.
Benchmarks can't be read under identical conditions
In Meta's comparison table, Muse Glimmer-30B scored 75.5 on MCP Atlas, 76.0 on SWE-Bench Verified, 65.9 on OSWorld-Verified, and 51.7 on TerminalBench 2.1. In the table alongside Gemma4-31B and Qwen3.6-27B, results vary by item. For example, on SWE-Bench Verified and OSWorld-Verified, Qwen3.6-27B's scores exceed Muse Glimmer's.
Meta's evaluation methodology notes that Muse Glimmer was measured with high reasoning strength, temperature 1.0, top_p 0.95, and top_k 64. Third-party models were evaluated best-effort under the same framework, but tools and system prompts were not optimized for each model. As such, the comparison table may not represent the best possible performance each model can achieve.
The impact of quantization also has a range. Meta reports an average accuracy drop of 0.2% for K-Quant-Dynamic and 1.0% for K-Quant-17GB across 15 benchmarks. This is an average under the company's own evaluation conditions and is not a guarantee that there will be no degradation across all agentic tasks or image inputs. For local deployment, verification would need to include the actual tools used, context length, and reversibility of operations.
Spark's API and Glimmer's weight release are separate paths
Muse Spark is a model that has been offered through Meta AI and meta.ai. At the time of its announcement in April 2026, it was a private API preview available to select users. In July, Meta released Muse Spark 1.1 as a public preview on the Meta Model API. This did not involve releasing Spark's own weights.
Glimmer, on the other hand, is a locally-oriented public model derived from Muse Spark, whose weights can be obtained and run. Meta's vision of "personal intelligence" also includes concepts like a personal agent running 24 hours a day, a free tier for billions of users, and a dynamic auction mechanism aimed at greater compute access. However, these represent visions articulated by Zuckerberg, and do not determine Muse Glimmer's product pricing or release timeline.
What's worth checking immediately after release is how far integration with llama.cpp, MLX, and ExecuTorch has progressed. Support for Ollama and LM Studio also remains only "planned" at the time of announcement. Meta's measurements alone cannot determine on-device agent operations. Whether the model can operate within a 24GB or 32GB environment while retaining the necessary context and tools will determine Glimmer's value.