Meta opens up Muse Spark Open, NVIDIA ships Nemotron 3.5 Super VL and Bolt, Grok Voice 3.0, and OpenTelemetry for the Claude API

Meta took the Muse Spark model it released in August, added more safety post-training, and published the weights. It's an MoE (mixture-of-experts: only some expert blocks fire for each input) with 816B total parameters and about 42B active per token. It accepts text, images, video and audio, and has a 1M-token context window. Weights are on Hugging Face as meta-models/Muse-Spark-Open-816B-A42B in BF16 and NVFP4, and SGLang, vLLM, TensorRT-LLM, MLX, llama.cpp and Unsloth support it from day one. With llama.cpp:

llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q4_K_S

In Meta's own table it scores 89.2 on MCP-AtlasVendor claim and 76.4 on Terminal-Bench 2.1Vendor claim. It beats Kimi K3 and GLM 5.3 on tool use but trails Kimi K3 (85.0) on Terminal-Bench, so try it first for tool-heavy business agents rather than pure coding. It runs on two M5 Ultra Mac Studios with 512GB each, or in roughly 300GB of memory with a 2-bit GGUF, so internal agents that can't send data off-site finally have a serious open option.

Super VL is a multimodal MoE with 120B parameters, 12B active, built for always-on agents that read web pages, PDFs, tables and charts and call tools. Bolt 74B-A9B uses native 2-bit quantization to fit in 24GB, so it can run local agents on an RTX PC or DGX Spark; a 4GB Bolt E4B follows on Oct 23. Weights are on Hugging Face, OpenRouter and NVIDIA NIM, and Bolt works with llama.cpp, Ollama and LM Studio.

llama-server -hf nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF

NVIDIA says it's up to 5x fasterVendor claim than other open models in its class, a big deal if your agents chew through piles of documents. The training datasets and recipes are public too, so post-training on your own data is easier, and NVIDIA says Atlassian, CrowdStrike and Perplexity are already customizing it. Bolt ranks first in its class on PinchBenchVendor claim, which makes it worth a try for work you'd rather keep on a laptop than send to the cloud.

A speech-to-speech model that hears audio and answers in audio, tuned to handle customer support calls end to end. It's available in the Voice Agent API and Voice Agent Builder, and agents pointed at grok-voice-latest moved to 3.0 automatically with no code changes.

It sits at #1 on the Artificial Analysis speech-to-speech index with 86.9Independent, and xAI reports 97% on Full Duplex BenchVendor claim, which measures interruptions and turn-taking. Responses start in under a secondVendor claim, and xAI says it handles 70% of Starlink support calls on its ownVendor claim. Teams looking at call-center or booking automation should test it next to GPT-Live-1 and Gemini Live.

This exports Messages API requests, tool use and Managed Agents sessions as OpenTelemetry spans (OpenTelemetry is the industry standard for shipping observability data). Add a destination in the Claude Console to get API-side spans; register a tracer in any of the seven SDK languages, including Python, TypeScript, Go and Java, to get SDK-side spans. By default only metadata such as token counts and latency goes out; turn on 'Full content' to include prompts and responses.

Any tool that accepts OTLP, such as Datadog, Grafana, Langfuse or Sentry, can now show Claude calls stitched into your service's own traces, so slow requests and failed tool calls show up on one screen. You can also build eval sets from real traffic. Delivery is best effort and spans can go missing, though, so use the Usage API for cost accounting.