Search “Meta”

Fireworks launches Meta's open-weight model Muse Spark Open

On October 20, 2026, Fireworks launched Meta's open-weight model Muse Spark Open on the day of its release, making it available to users through its serverless API. The model is an 816B-parameter MoE model that activates 42B parameters per token, has a 1M-token context window, and is aimed at long-horizon coding and agentic tasks. Fireworks said the model ranks first among 75 models on Vals AI's Harvey legal agent benchmark with 25.4%, nearly four times that of Claude Opus 5 (6.7%); on the Rogo Big Finance Bench, it surpasses Claude Opus 5.5 at every difficulty level. Fireworks also said it validated the weights with its own production stack, scoring 91.1% on AIME26, 89.7% on GPQA Diamond, and 84.3% on SWE-bench Verified, with consistent results when rerun using SGLang. On Fireworks, the model has zero data retention enabled by default, uses only compute located in the United States with no additional surcharge, and supports fine-tuning with serverless LoRA.

View story:Fireworks Launches Meta's Open-Weight Model Muse Spark Open

English translation is being prepared.

Meta is developing Confidential VMs functionality on the Muse web client, allowing users to run Muse tasks in an encrypted confidential environment. Meta previously previewed that this feature would launch before the end of 2026. A newly surfaced hidden setting shows that the feature will initially be available only to Meta AI subscribers; after toggling it on, voice mode, image and video generation, and wearable device support will be temporarily unavailable. Meta says conversations run in a secure encrypted environment and that Muse data will not be used to improve Meta's AI. The feature is currently still in development, and there is no further evidence of an official launch date.

View story:Meta is developing Confidential VMs functionality on the Muse web client

Meta says it is training the Muse foundation model on the Prometheus supercomputing cluster

Meta announced that the first phase of its gigawatt-scale supercomputing cluster Prometheus has been completed in New Albany, Ohio, and is being used to train the Muse series of foundation models, including Muse Wonder. The cluster currently consists of hundreds of thousands of GPUs, with a full-scale target of about 1 million, and has deployed two generations of platforms, NVIDIA GB200 NVL72 (Catalina) and GB300 NVL72 (Clemente), with Vera Rubin NVL72 to be introduced later. Meta says Muse Wonder's pre-training goodput exceeds 85%, reinforcement learning goodput exceeds 90%, and the interruption rate has been reduced by 99% from its mid-2025 peak. Subsequent reports added that during Muse Wonder training, an interruption occurred on average only about every 47K GPU hours; Meta said that in a cluster of this scale, even a single host failing to meet performance standards would slow down an entire training round, so Prometheus was designed for large-scale fault tolerance.

English translation is being prepared.

Meta announced that the first phase of its gigawatt-scale supercomputing cluster Prometheus has been completed in New Albany, Ohio, and is being used to train the Muse series of foundation models, including Muse Wonder. The cluster currently consists of hundreds of thousands of GPUs, with a full-scale target of about 1 million, and has deployed two generations of platforms, NVIDIA GB200 NVL72 (Catalina) and GB300 NVL72 (Clemente), with Vera Rubin NVL72 to be introduced later. Meta says Muse Wonder's pre-training goodput exceeds 85%, reinforcement learning goodput exceeds 90%, and the interruption rate has been reduced by 99% from its mid-2025 peak. Subsequent reports added that during Muse Wonder training, an interruption occurred on average only about every 47K GPU hours; Meta said that in a cluster of this scale, even a single host failing to meet performance standards would slow down an entire training round, so Prometheus was designed for large-scale fault tolerance.

View story:Meta says it is training the Muse foundation model on the Prometheus supercomputing cluster

Meta compresses its agent model Muse Glimmer 30B to run on-device

On October 20, 2026, Meta released the technical report for Muse Glimmer, a 30B open-weight model aimed at agent tasks on consumer-grade hardware, running at 4-bit precision on a single consumer GPU. The reporter adds that the model leads open-source models of the same scale in tool calling and multi-step planning, surpassing Qwen3.6-27B and Gemma4-31B on some metrics; after 4-bit quantization it is about 17GB, targeting devices with 24GB of VRAM, and uses DFlash speculative decoding to reach about 233 tokens/s on an RTX 5090, roughly 3.1 times faster than ordinary decoding. There are currently no subsequent developments.

View story:Meta compresses its Agent model Muse Glimmer 30B to run on-device

English translation is being prepared.

Meta updated its commercial terms, removing language that had been interpreted as allowing it to scrape website content to train AI. Neiman Lab and Cornell University law professor James Grimmelmann previously argued that the clause was intended to let Meta use scraped content for free to train AI, and could bypass blocking instructions in robots.txt. Meta Vice President of Communications Andy Stone and spokesperson Daniel Roberts had responded that Meta's use of web crawlers had not changed, and this terms update directly removed the disputed language. Other reports said that when Meta updated its business services terms, it had stated it had the right to use content from business account domains, web pages or URLs to train AI models, and deleted that language after controversy; Meta said the update was merely a consolidation of existing advertising terms, not an expansion of crawler behavior, and that the revised terms limit crawlers and data collection to advertising and business products only, while site owners can still control crawler permissions via robots.txt.

View story:Meta updates its commercial terms, removing language that could have allowed scraping of news website content

English translation is being prepared.

English translation is being prepared.

View story:English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

View story:English translation is being prepared.

Meta releases open-weight model Muse Spark Open

On October 15, 2026, Meta released the open-weight model Muse Spark Open, with the weights going live on Hugging Face the same day under the Muse Model License. According to the PyTorch Blog, the model has 816B total parameters, 42B activated parameters, and a 1M token context window; it is natively multimodal and aimed at coding and agentic tasks, adds safety post-training on top of Muse Spark, and can run on SGLang and vLLM. On October 16, the model launched on Amazon SageMaker JumpStart with two variants, BF16 and NVFP4; the same day it entered live testing in Arena's Agent Arena and launched in Code Arena's WebDev and other projects. On October 17, Arena announced that Muse Spark Open (Max) ranked first among US open models on the Code Arena WebDev leaderboard with a score of 1511, placing 13th among all open models.

Meta Releases Muse Glimmer Technical Report

Meta has released the Muse Glimmer technical report. The report introduces a 30B open-weight model aimed at agentic tasks on consumer-grade hardware. According to the report, at 4-bit precision the model can run at practical speeds on a single consumer GPU in a Mac or PC, and it leads open-source models of the same scale in tool use and multi-step planning. The report covers aspects including model architecture, distillation recipe, and quantization. There are no further developments at present.

Meta releases Muse Glimmer technical report

Meta has released the Muse Glimmer technical report. The report introduces a 30B open-weight model aimed at agentic tasks on consumer-grade hardware. According to the report, at 4-bit precision the model can run at practical speeds on a single consumer GPU in a Mac or PC, and it leads open-source models of the same scale in tool use and multi-step planning. The report covers aspects including model architecture, distillation recipe, and quantization. There are no further developments at present.

View story:Meta Releases Muse Glimmer Technical Report

Meta compresses its Agent model Muse Glimmer 30B to run on-device

On October 20, 2026, Meta released the technical report for Muse Glimmer, a 30B open-weight model aimed at agent tasks on consumer-grade hardware, running at 4-bit precision on a single consumer GPU. The reporter adds that the model leads open-source models of the same scale in tool calling and multi-step planning, surpassing Qwen3.6-27B and Gemma4-31B on some metrics; after 4-bit quantization it is about 17GB, targeting devices with 24GB of VRAM, and uses DFlash speculative decoding to reach about 233 tokens/s on an RTX 5090, roughly 3.1 times faster than ordinary decoding. There are currently no subsequent developments.

Meta releases technical details of the Prometheus supercluster

Meta's engineering team has published technical details of the Prometheus supercluster. Located in New Albany, Ohio, the cluster currently consists of hundreds of thousands of GPUs and operates as a single machine, with plans to scale to approximately 1 million GPUs. Meta confirmed that the first phase has already used the cluster to train the Muse family of foundation models (including Muse Wonder), with Muse Wonder achieving pretraining goodput of over 85% and reinforcement learning goodput of over 90%. Meta also disclosed two platforms: Catalina, based on NVIDIA GB200 NVL72, and Clemente, based on GB300 NVL72, as well as the upcoming Vera Rubin NVL72 platform.

Meta updates its commercial terms, removing language that could have allowed scraping of news website content

Meta updated its commercial terms, removing language that had been interpreted as allowing it to scrape website content to train AI. Neiman Lab and Cornell University law professor James Grimmelmann previously argued that the clause was intended to let Meta use scraped content for free to train AI, and could bypass blocking instructions in robots.txt. Meta Vice President of Communications Andy Stone and spokesperson Daniel Roberts had responded that Meta's use of web crawlers had not changed, and this terms update directly removed the disputed language. Other reports said that when Meta updated its business services terms, it had stated it had the right to use content from business account domains, web pages or URLs to train AI models, and deleted that language after controversy; Meta said the update was merely a consolidation of existing advertising terms, not an expansion of crawler behavior, and that the revised terms limit crawlers and data collection to advertising and business products only, while site owners can still control crawler permissions via robots.txt.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

View story:English translation is being prepared.

Meta releases technical details of Prometheus supercluster

Meta's engineering team has published technical details of the Prometheus supercluster. Located in New Albany, Ohio, the cluster currently consists of hundreds of thousands of GPUs and operates as a single machine, with plans to scale to approximately 1 million GPUs. Meta confirmed that the first phase has already used the cluster to train the Muse family of foundation models (including Muse Wonder), with Muse Wonder achieving pretraining goodput of over 85% and reinforcement learning goodput of over 90%. Meta also disclosed two platforms: Catalina, based on NVIDIA GB200 NVL72, and Clemente, based on GB300 NVL72, as well as the upcoming Vera Rubin NVL72 platform.

View story:Meta releases technical details of the Prometheus supercluster