AI Daily · Oct 16, 2026: Meta releases Muse Spark Open, an 816B open-weight model
Meta released Muse Spark Open, an 816B open-weight model with 1M context and full multimodality at $1.25/$4.25 per M tokens; StepFun open-sourced the 600B sparse MoE Step 5 Preview, priced $1/$2.70 on DeepInfra.
Showing the latest briefing (Oct 16, 2026). Every day is generated the next morning.
Pick a date
-
Meta releases Muse Spark Open, 816B open-weight model
Muse Spark Open has 816B total / 42B active parameters, a 1M-token context window, and native multimodality across text, image, video and audio. Weights are downloadable via Hugging Face, SageMaker, Google Cloud Model Garden and Microsoft Foundry, with API access on Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI under the Muse Model License; it runs on two Apple M5 Ultra Mac S
-
Meta's Muse Spark Open arrives on OpenRouter with 1M context
OpenRouter lists Muse Spark Open, described as the strongest US open-weight model, built for agentic coding with 1M context and image, video, audio and PDF input, priced at $1.25/M input and $4.25/M output; it also launched day 0 on Baseten, Fireworks, Nebius and Together, with GGUF quantizations available.
-
Step 5 Preview goes live on DeepInfra at day zero
StepFun's Step 5 Preview, a 600B open-weight MoE with 27B active parameters, 1M context and text-plus-vision in FP8, is live on DeepInfra at $1.00 input, $2.70 output and $0.05 cached per million tokens.
-
Step 5 Preview open weights: 600B sparse MoE, 1M context
StepFun released Step 5 Preview open weights in BF16 and FP8 on Hugging Face and ModelScope with a technical blog; it is a sparse MoE with 600B total and 27B active parameters per token and 1M-token context, and an experiment had 6 open models write the same M&A memo for 80 companies judged by two AI judges.
-
NVIDIA Nemotron 3.5 Super VL released with 120B MoE and multimodal support
NVIDIA released Nemotron 3.5 Super VL, a 120B parameter MoE model with 12B active parameters, supporting text, image, document, GUI, and video input. It uses a hybrid Mamba-Transformer architecture with MTP and NVFP4 KV cache, and is available for free in Cline.
-
Perplexity Open Sources SPACE Agent Sandbox Under Apache 2.0
Perplexity open sourced SPACE, the sandbox platform behind Perplexity Computer, under Apache 2.0. Each agent task gets its own Firecracker microVM with credentials and outbound traffic controlled outside the VM; in testing, nine model configurations with root inside the VM made 108 escape attempts and none reached the host.
-
Tencent open-sources Hy Memory 2.0 model, engine, and MemHub app
Tencent released Hy Memory 2.0, an open-source memory stack: the Hy-Memory-30B-A3B model (3B active parameters, 24GB+ VRAM recommended), the Hy Memory engine SDK, and the MemHub CLI/Mac app that syncs across 10+ AI tools. It reports first place on four benchmarks, leading MemoryArena by 10.2 points while cutting memory volume 51.8% and recall tokens 39.3%, and claims the 3B-active model beats Opus
-
Reflection AI releases Beam open weights under Apache 2.0
Beam is now available across platforms with open weights under Apache 2.0, released with 40+ partners along with tokenizer, reference implementation, BF16/FP8/NVFP4 weights, a DFlash2 drafter and the Mirror CLI harness.
-
OpenRouter launches Armor beta to block out-of-policy tool calls
OpenRouter's Armor beta uses TypeSafe AI's Jev model to block out-of-policy tool calls before execution, claiming higher accuracy and lower cost than Claude Code's auto classifier and Sonnet 5 on an internal test set; agents opt in via an Armor plugin.
-
Meta Muse Wonder lands in Cursor at $4/$20 per million tokens
Muse Wonder from Meta Superintelligence Labs is now available in Cursor, scoring 46.1% on CursorBench at default effort, the highest of any Muse model, priced at $4/M input and $20/M output tokens.