On August 10, Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter model engineered to run agent workflows on a single consumer GPU. Capable agent models have required data center hardware and per-token API billing. Glimmer compresses a frontier-derived model into 24 gigabytes of VRAM, pairs it with speculative decoding for interactive speed, and wires it into the Model Context Protocol for tool calling. The model occupies a deliberate third lane in Meta’s AI lineup, alongside the open-frontier Llama series and the closed Muse Spark flagship, aimed at always-on local deployment where queries never leave the device.

Compression arithmetic

The lab distilled Glimmer from Muse Spark 1.2, Meta’s closed flagship, through logit distillation, a training method in which the smaller model learned to match the output distributions of the larger one. This approach preserves more of the teacher model’s behavior than training on generated text alone. Glimmer is Meta’s first open-weight release since Llama 4, and the Apache 2.0 license removes the monthly-active-user cap that restricted commercial use of earlier Meta models.

Training proceeded in three phases: logit distillation from Spark, a second phase emphasizing long-context and agent-heavy data, and a final stage combining supervised fine-tuning with on-policy distillation and reinforcement learning. The pipeline prioritized agentic capabilities such as tool calling, multi-step planning, and failure recovery.

At full floating-point precision, 30 billion parameters require over 55 gigabytes of memory, well beyond any consumer graphics card. Meta’s 4-bit quantization reduces that footprint to under 20 gigabytes, leaving room for the KV cache, a 1.8-billion-parameter vision encoder, and the speculative decoding drafter within a 24- or 32-gigabyte VRAM envelope. Meta reports less than 1% accuracy degradation at the smaller quantization level.

Protocols at speed

DFlash, a companion drafter network that ships with Glimmer, proposes entire blocks of tokens for the main model to verify in parallel. This approach replaces the standard one-token-at-a-time generation loop and yields measurable speedups: 3.1x on an NVIDIA RTX 5090, 1.8x on an Apple M5 Max, and 1.5x on an M4 Max, with no reported quality loss. At 233 tokens per second on the RTX 5090, local inference reaches interactive speed.

Glimmer supports the Model Context Protocol natively, which allows it to call external tools, manage state across multi-step workflows, and retry failed invocations without external orchestration. A 1.8-billion-parameter ViT-G/14 vision encoder accepts mixed text and image input, enabling screenshot and document comprehension within agent workflows. The model also offers adjustable reasoning effort from low to extra-high, so developers can trade quality for speed depending on the task.

Ollama, LM Studio, Together AI, Fireworks AI, and OpenRouter supported the model on release day, with optimized builds for llama.cpp, MLX, and ExecuTorch following within days. Hardware optimization partnerships with AMD, Arm, Dell, Intel, and NVIDIA accompanied the launch, a breadth of support suggesting that Meta intends Glimmer as production infrastructure.

Split benchmarks

Meta’s own benchmarks place Glimmer ahead of Gemma4-31B and Qwen3.6-27B on agentic tool-use tasks. The model scored 75.5 on MCP Atlas, a benchmark for multi-step tool calling, compared to 62.5 for Qwen and 54.2 for Gemma. Similar margins appeared on DeepSearch QA (74.6 versus 71.1 and 61.7) and τ³-Banking (23.5 versus 16.7 and 15.1). Qwen retained the lead on sustained coding and computer use: 77.2 to 76.0 on SWE-Bench Verified, 75.6 to 65.9 on OSWorld, and 60.7 to 51.7 on TerminalBench 2.1. All of these figures are self-reported by Meta, which ran the benchmarks and selected the comparison set.

Artificial Analysis provided a less favorable independent assessment, scoring Glimmer at 953 Elo on GDPval-AA v2, a benchmark for realistic knowledge-work tasks, below the 1,000-point human baseline. The evaluator recorded an 82% hallucination rate on AA-Omniscience, compared to 49% for Qwen3.6-27B. Glimmer tended to generate confident answers even when it lacked the relevant knowledge, a failure mode that the structured tool-use benchmarks did not capture.

These results trace a deployment boundary. Glimmer handles structured agent workflows well enough for supervised use, while the hallucination rate rules out unsupervised autonomous operation on tasks in which errors carry consequences.

Proof of concept

Glimmer’s compression pipeline may prove more consequential than the model itself. The combination of logit distillation from a frontier teacher, aggressive quantization, and speculative decoding demonstrated a repeatable method for moving cloud-only capabilities to local hardware. Zuckerberg promised alongside the release to open the weights for Muse Spark 1.2, which would push a frontier-class model through a similar process. If subsequent iterations close the hallucination gap, the economics of agent deployment shift: local inference carries a one-time hardware cost where cloud inference carries a per-token cost that compounds with every agent action.

Subscribe now

AI Central

Agents Leave the Cloud