THE FORWARD PASS
Advertise
Archive

Daily edition October 16, 2026

Daily AI-generated issue

AI engineering, October 16, 2026

Meta launches Muse Spark Open, NVIDIA launches Nemotron Super VL and Bolt 74B-A9B, Anthropic opens Opus 5.5 fast mode on Claude, Mistral releases 209M-parameter Timestral

This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent.

  • Agents
  • APIs
  • Audio
  • Benchmarks
  • Business
Edition
Daily
Published
October 16, 2026
Read time
3 min
Stories
3

NVIDIA launched Nemotron Super VL and Bolt 74B-A9B. Bolt 74B-A9B fits 74B parameters into 24GB of memory using native 2-bit quantization.

Meanwhile, Meta launched Muse Spark Open with downloadable weights and hosted inference through seven providers. Anthropic opened Opus 5.5 fast mode on the Claude platform at $8/$40 per million tokens, with no waitlist.

If you read one thing today: Muse Spark Open. Hosted APIs let you evaluate the model before committing hardware to a vLLM or SGLang deployment.

Key takeaways

  1. 01The 816B-parameter MoE activates 42B parameters per token.
  2. 02Text, image, video, and audio inputs share a 1M-token context window.
  3. 03BF16 weights take about 1.6TB, while NVFP4 takes about 530GB.
  4. 04Serving options include SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth.

Top News · 713 HN points

Muse Spark now has an open-weight version you can download or access through hosted inference. Meta launched Muse Spark Open with weights on Hugging Face and hosted options from seven providers, so you can evaluate it without first setting up a self-hosted deployment.

Here's what ships:

  • The 816B-parameter MoE activates 42B parameters per token.
  • Text, image, video, and audio inputs share a 1M-token context window.
  • BF16 weights take about 1.6TB, while NVFP4 takes about 530GB.
  • Serving options include SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth.

One catch: the Muse Model License has limited commercial terms. Hosted access also varies by provider, with Groq in limited preview.

Try it: Download the Hugging Face weights, or use Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, or Together AI for hosted inference.

Sources: research.meta.ai · dev.meta.ai

Top News · Official · nvidia.com

NVIDIA launched Nemotron Super VL for multimodal agents and Bolt 74B-A9B for local deployment. NVIDIA reports up to 5× faster throughput for Super VL than other models in its class on the Artificial Analysis Intelligence Index leaderboard.

Here's what ships:

  • Super VL is available now on Hugging Face, Open Router, ModelScope community, and as NVIDIA NIM.
  • Super VL has 120B parameters and handles tool use, multistep workflows, and visual documents.
  • Bolt 74B-A9B is available now on Hugging Face and uses native 2-bit quantization to fit in 24GB.
  • Bolt supports llama.cpp, Ollama, LM Studio, Unsloth, and vLLM.
  • Super VL includes open weights, datasets, and training recipes for post-training on your own data.

One catch: the 4GB Bolt E4B configuration is not available yet. NVIDIA says it arrives Oct. 23.

Sources: blogs.nvidia.com

Top News · Reported · aihot.news · translated from Chinese

If you have been waiting for access to Opus 5.5 fast mode on the Claude platform, you no longer need a waitlist slot. Anthropic has ended the research preview, opening access to the faster serving option for the same Opus 5.5 model.

Here's what you get:

  • Fast mode is priced at $8/$40 per million tokens on the Claude platform.
  • Opus 5.5 runs at up to 2.5 times the speed in fast mode, without switching models.
  • Microsoft Foundry also offers fast mode, with inference hosted on Azure.

One catch: Google Cloud support is still coming soon. Claude Code fast mode remains unchanged and continues to use additional usage credits, so this update does not change your Claude Code setup.

Sources: aihot.news

Also in this issue

Signals

  1. FLUX 3 Image Klein generates and edits images on 8GB NVIDIA RTX GPUs
  2. AMD Instinct MI355X reportedly beats NVIDIA B300 by 24.4% in Kimi K3 AgentX throughput per GPU · Unconfirmed · vllm.ai
  3. StreamErase removes video objects at up to 26 FPS on a single GPU · 3 Upvotes
  4. Mistral releases Timestral, a 209M-parameter model topping three time-series forecasting benchmarks
  5. Hugging Face Transformers adds Meta’s Muse Spark Open with up to a 1,048,576-token multimodal context · 3 Stars today
  6. NVIDIA releases 4.38B-parameter DFlash 2 draft model for Muse Spark Open speculative decoding

Keep reading