Meta 发布并开源 Muse Glimmer

2026 年 8 月 10 日，Meta 超级智能实验室发布 Muse Glimmer：约 30B 参数的稠密多模态智能体模型，Apache 2.0 开源，可在 24GB 显存消费级 GPU 或 M4/M5 Max MacBook 上完全本地运行。同一天，扎克伯格发表长篇檄文《The Future Is for Everyone》，批评「限制开源」路线，宣布 Meta 重返开源。

先看模型。Muse Glimmer 由闭源基座 Muse Spark 逻辑蒸馏而来——这是理解它的关键：它不是独立训练的基座，而是「前沿闭源基座 + 中型开源专用模型」双线策略的一环。上下文 128K，提供两套 4bit 量化（K Quant 约 17GB、Dynamic K Quant 约 20GB），配套 DFlash 区块并行推测解码，RTX 5090 上生成速度从 74.9 提升到 233.4 token/s。官方口径里，它在 MCP Atlas、DeepSearch QA、Gaia2 等多个 Agent 基准上优于 Gemma4-31B 与 Qwen3.6-27B——需要带着厂商自述的边界读，但「30B 稠密模型可以本地跑 Agent 工作流」这件事本身是确定的。

再看路线。Meta 在 Llama 系列之后转向闭源旗舰 Muse Spark，开源社区对其路线反复摇摆多有批评。扎克伯格这篇檄文把「限制开源」直接点名批评，并预告 Muse Spark 1.2 的开放权重版本——「闭源基座 + 开源中型」不是权宜之计，而是被写进战略的组合打法。对社区来说，这篇檄文比模型本身更说明问题：Meta 需要开发者生态，而开源是拿回生态的入场券。

端侧 Agent 的竞争格局同样值得读。Gemma 4、Qwen3.6-27B 等中型开源模型此前主导了本地 Agent 场景；Muse Glimmer 把能力门槛抬到 30B 档，同时用「完全本地运行」回应隐私与成本诉求。对开发者，这是一个可完全本地运行、Apache 2.0 的 30B 智能体模型，数据不出设备、成本可控；对行业，端侧 Agent 模型的能力门槛被抬到 30B 档，「闭源基座 + 开源中型」成为新的组合打法。

Muse Glimmer 的意义不在单个数字。它把 Meta 的开源路线从「Llama 时代」重新接上，用「闭源基座蒸馏 + 开源中型」的组合回应社区；同时把端侧 Agent 模型的能力门槛抬到 30B 档，本地运行成为可交付的产品形态。至于「重返开源」的说法——Meta 从未离开过开源，只是离开过社区。这一次，它带着 30B 的模型和一篇檄文回来了。

On August 10, 2026, Meta Superintelligence Labs released Muse Glimmer: a ~30B dense multimodal agent model, open under Apache 2.0, fully runnable on a 24GB consumer GPU or M4/M5 Max MacBook. The same day, Zuckerberg published a long manifesto, "The Future Is for Everyone," criticizing the restrict-open-source path and declaring Meta's return to open source.

On the model: Muse Glimmer is distilled from the closed Muse Spark base via logit distillation—this is the key to understanding it. It is not an independently trained base but part of a "frontier closed base plus mid-size open model" dual-track strategy. With a 128K context, two 4-bit quantizations ship (K Quant ~17GB, Dynamic K Quant ~20GB), plus DFlash block-parallel speculative decoding lifting RTX 5090 generation from 74.9 to 233.4 tokens/s. By Meta's account it beats Gemma4-31B and Qwen3.6-27B on MCP Atlas, DeepSearch QA, and Gaia2—read with the vendor-self-reported caveat, but "a 30B dense model can run agent workflows locally" is itself a fact.

On the course: after the Llama line, Meta turned to the closed Muse Spark flagship, drawing community criticism over its wavering. Zuckerberg's manifesto names the restrict-open-source path directly and teases open weights for Muse Spark 1.2—"closed base plus open mid-size" is not a stopgap but a strategy written into the playbook. For the community, the manifesto says more than the model: Meta needs the developer ecosystem, and open source is the ticket back in.

The on-device agent landscape deserves attention too. Gemma 4, Qwen3.6-27B, and other mid-size open models had dominated local agent scenarios; Muse Glimmer raised the capability bar to the 30B tier while answering privacy and cost demands with "fully local." For developers, this is a fully local, Apache 2.0 30B agent model—data stays on device, cost under control; for the industry, the on-device agent bar rose to the 30B tier, with closed-base-plus-open-mid-size as the new playbook.

Muse Glimmer's meaning is not in any single number. It reconnected Meta's open-source course after the Llama era with a closed-base-distilled, open-mid-size combination; it also raised the on-device agent bar to the 30B tier, making local operation a deliverable product form. As for "return to open source"—Meta never left open source; it left the community. This time it came back with a 30B model and a manifesto.