Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context
https://www.marktechpost.com/2026/10/15/meta-muse-spark-open-816b-open-weight-model/📌 【Meta 最新開源】816B 的 Muse Spark,兩臺 Mac Studio 就能跑
TL;DR:Meta 釋出 816B 參數開源多模態 MoE 模型,支援文字/影像/影片/音訊輸入與 1M context,量化後可在兩臺 Mac Studio 運行。
816B 參數的模型通常意味著需要一整個資料中心才能服務,但 Meta 這次把量化版本的部署門檻壓到了兩臺 Apple M5 Ultra Mac Studio——這對「開源旗艦模型」的想像是個不小的轉變。
🤔 Meta 的開源旗艦
Meta Superintelligence Labs 釋出了 Muse Spark Open,是其 Muse Spark 模型的開源權重版本。Meta 團隊稱其為目前最強的美國開源權重模型。這是一個原生多模態的 MoE 模型,鎖定程式碼與長流程 agentic 工作,輸入涵蓋文字、影像、影片與音訊,輸出為文字,context window 達 1M tokens。
🧩 816B 總參數、42B 啟動的架構
模型卡列出 62 層、hidden dimension 為 8,192。每個 MoE 層有 256 個專家,top-k 路由每個 token 啟動 8 個。Attention 採分組查詢注意力,128 個 query head 對 16 個 KV head。視覺輸入由獨立的 18 億參數 Perception Encoder 處理。詞彙表達 202,048 個 token,訓練資料涵蓋超過 100 種語言。推理深度可透過 reasoning_effort 參數設定為 low、medium、high 或 xhigh,預設為 xhigh。開源版本額外做了安全後訓練,Meta 團隊表示這讓模型在沒有系統層級防護機制的情況下也更穩健。
📊 美國最強,但編碼仍落後中國開源模型
Meta 在最高推理強度下,將 Muse Spark Open 與 Kimi K3、GLM 5.3、Qwen3.8-2.4T-A95B 等模型比較。編碼是它的弱項:在 Terminal-Bench 2.1(搭配 Terminus 2 harness)上拿下 76.4 分,落後 Kimi K3 的 85.0、GLM 5.3 的 83.9 與 Qwen3.8 的 82.0。在 Artificial Analysis Intelligence Index 上,Muse Spark Open 得分 39,同樣落後 GLM 5.3(45)、Kimi K3(44)與 Qwen3.8(40)。換句話說,「最強美國開源模型」的說法站得住腳,但中國開源模型在整體能力上仍然領先。部分對手分數為自行回報,評測方法論有列出各數據來源。
💡 量化讓 816B 變得可部署
根據 Meta 的開發者指南,BF16 權重約 1.6TB,NVFP4 只量化專家部分,壓到約 530GB。模型卡建議 NVFP4 至少需要 700GB VRAM,BF16 則需要 2TB。Day-one 支援引擎包括 SGLang、vLLM、TensorRT-LLM、Hugging Face Transformers、MLX、llama.cpp 與 Unsloth。Hosted API 已在 Baseten、Fireworks AI、Groq、Modal、Nebius、OpenRouter、Together AI 上線,OpenRouter 標價為每百萬輸入 token 1.25 美元、每百萬輸出 token 4.25 美元。
兩個 draft 模型用於加速解碼:DFlash V2 搭配 SGLang,DSpark 搭配 vLLM。Meta 回報在 AMD MI355X 上,DFlash V2 帶來 3.7 倍的單請求吞吐量提升。微調方面支援 TorchTitan、Hugging Face TRL 與 NVIDIA NeMo AutoModel。授權採用 Muse Model License,允許使用、微調、蒸餾與重新散布;其中內部使用、研究用途以及用於自家產品功能皆豁免於營收條款之外。
⚠️ 編碼能力是明顯短板
雖然整體 Intelligence Index 與多數評測上標榜「最強美國開源」,但在編碼這個對 agentic 工作極為關鍵的面向上,Muse Spark Open 明顯落後 Kimi K3、GLM 5.3 與 Qwen3.8,且部分對手分數屬自行回報,需謹慎解讀跨廠商比較。
🎯 實務啟示
如果你的場景需要處理音訊、影片等多模態輸入並要求長 context,Muse Spark Open 的 1M context 與原生多模態能力值得評估;但若核心需求是程式碼生成與 agentic coding,目前的評測顯示它不是這個面向的最佳選擇。NVFP4 把 816B 模型壓到兩臺 Mac Studio 可跑的規模,也為本地高階部署開了一條新路。
🔗 來源
- 標題:Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context
- 作者／機構:Asif Razzaq,MarkTechPost
- 連結:https://www.marktechpost.com/2026/10/15/meta-muse-spark-open-816b-open-weight-model/
#Meta #MuseSpark #OpenWeightLLM #MixtureOfExperts #Multimodal #LongContext #AgenticAI #NVFP4 #OpenSource #AIInfrastructure
原始資料 MarkTechPost · 收集於 2026-10-16
摘要原文
Meta Superintelligence Labs has released Muse Spark Open , the open-weight version of its Muse Spark model. Muse Spark Open is an 816B-parameter Mixture-of-Experts model that activates 42B parameters per token. It takes text, image, video and audio input, returns text, and holds a 1M-token context window. Meta team claims it to be the most capable US open-weight model to date. For developers, the real news is deployment: the NVFP4 build runs on 2 Apple M5 Ultra Mac Studios. Muse Spark Open is a natively multimodal MoE model from Meta, built for coding and long-horizon agentic work. Weights ship in BF16 and NVFP4 on Hugging Face. The open version received extra safety post-training. Meta team states that this makes it more robust without system-level guardrails. The model card lists 62 layers and a hidden dimension of 8192. Each MoE layer holds 256 experts, and top-k routing activates 8 per token. Attention is grouped-query, with 128 query heads and 16 KV heads. A separate 1.8B-parameter Perception Encoder handles visual input. The vocabulary has 202,048 tokens, and training data covers more than 100 languages. Reasoning depth is set with reasoning_effort : low, medium, high or xhigh. The default is xhigh. Meta compares it with Kimi K3, GLM 5.3, Qwen3.8-2.4T-A95B and others at maximum reasoning effort. Its strongest results are on tool use and knowledge reliability: Coding is the weak spot. On Terminal-Bench 2.1 with the Terminus 2 harness, it scores 76.4. Kimi K3 reaches 85.0, GLM 5.3 scores 83.9 and Qwen3.8 scores 82.0. Its Artificial Analysis Intelligence Index is 39. That sits behind GLM 5.3 (45), Kimi K3 (44) and Qwen3.8 (40). So “most capable US” holds, but Chinese open models still lead overall. Some rival scores are self-reported; the evaluation methodology lists each source. Per Meta’s developer guide , BF16 weights are about 1.6 TB. NVFP4 quantizes only the experts and lands near 530 GB. The model card recommends at least 700 GB of VRAM for NVFP4 and 2 TB for BF16. Day-one engines include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth. Hosted APIs are live on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. OpenRouter lists $1.25 per 1M input tokens and $4.25 per 1M output tokens. Two draft models speed up decoding: DFlash V2 for SGLang and DSpark for vLLM. Meta reports 3.7x single-request throughput with DFlash V2 on AMD MI355X. Fine-tuning works with TorchTitan, Hugging Face TRL and NVIDIA NeMo AutoModel. The Muse Model License permits use, fine-tuning, distillation and redistribution. 2 conditions matter most: Internal use, research and powering your own product features are exempt from the revenue clause. Check out the model weights , research blog , developer guide and evaluation methodology . All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter . Wait! are you on telegram? now you can join us on telegram as well. The post Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成