# Muse Spark Open, FLUX.3 Klein, Timestral, Nemotron 3.5 Super VL

> Meta opens an 816B multimodal MoE, Black Forest Labs ships a 5.6B image model, Mistral releases Timestral, and NVIDIA and IBM add open agent models.

Published 15 Oct 2026. Coverage: 14 Oct 2026 to 15 Oct 2026. By Ali Asif.

## Items

1. **Welcome Muse Spark Open** (Meta, Open weights): An open-weight multimodal mixture-of-experts model with about 816B total and 42B active parameters, a 1M token context window and a focus on agentic coding. [Primary source](https://huggingface.co/blog/meta-muse-spark-open)
2. **FLUX.3 Image Klein** (Black Forest Labs, Image, video, world): A 5.6B open-weight image model for text-to-image, editing and up to 10 reference images, with an INT8 build that runs on GPUs with 8 GB of VRAM. [Primary source](https://bfl.ai/blog/flux-3-image-klein)
3. **Introducing Timestral** (Mistral, Open weights): Mistral's first time-series foundation models, from 5M to 209M parameters, with open weights and CPU-only inference. [Primary source](https://mistral.ai/news/timestral)
4. **NVIDIA Nemotron 3.5 Super VL** (NVIDIA, Open weights): An open 120B multimodal MoE with 12B active parameters, scoring 80.8% on CharXiv and 63.44% on Hermes Bench. [Primary source](https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning)
5. **Granite Switch 4.1** (IBM, Open weights): An 8B and a 30B Granite family that adds decision, safety guardian and speech-to-text adapters to one endpoint at roughly the VRAM of the base model. [Primary source](https://huggingface.co/blog/ibm-granite/granite-switch-4-1)
6. **Accelerating the search for antimicrobial peptides with Google Antigravity** (Google DeepMind, Science): An agent filtered UniProt to a 97 amino acid bacteriocin candidate, checked novelty and structure, and drafted a wet-lab plan; validation is still under way. [Primary source](https://deepmind.google/blog/accelerating-antimicrobial-discovery)

A heavy day for open weights. Meta, NVIDIA, IBM, Mistral and Black Forest Labs all shipped models you can download, and DeepMind published an agent case study in biology.

## Meta opens Muse Spark

Muse Spark Open is a multimodal mixture-of-experts model with about 816B total parameters and about 42B active per token. It uses 256 experts per layer with 8 selected per token, and 62 layers in the language model. It takes text, images, audio and video, and outputs text only.

The context window reaches 1M tokens. Checkpoints come in BF16 and NVFP4, and Meta trained it for agentic coding, including tool calling. Reasoning cannot be switched off, but effort can be set to low, medium, high or xhigh. You can run it through Transformers, llama.cpp, SGLang, vLLM and MLX. The Hugging Face post reports no benchmark scores and does not state a license, so I would read the model card before building on it.

## FLUX.3 Image Klein

Black Forest Labs released a 5.6B open-weight image model that handles text-to-image, editing and up to 10 reference images, with native 4K output. By their own internal evaluations it gains 91 Elo in text-to-image and 70 in editing over FLUX.2 klein 9B, with 38% fewer parameters. The INT8 build runs on GeForce cards with 8 GB of VRAM or more, and the distilled model takes about 4 seconds for 1 MP on an RTX 5090. The license is the FLUX Kommunity License, free for commercial use under $5M in annual revenue.

## Timestral from Mistral

Timestral is Mistral's first time-series family, at 5M, 28M, 123M and 209M parameters. Mistral says the 209M model ranks first among pre-trained zero-shot models on GIFT-Eval, TIME and fev-bench, and that the 28M model beats Chronos-2 on two of them. It needs no GPU: a one-day-ahead forecast takes 10 ms on the 5M model and 140 ms on the 209M. Weights are open under a Modified MIT license, and a technical report is still to come.

## Nemotron 3.5 Super VL and Granite Switch 4.1

NVIDIA's Nemotron 3.5 Super VL is a 120B model with 12B active, distilled from Nemotron 3 Ultra. NVIDIA reports 80.8% on CharXiv and 63.44% on Hermes Bench, against 56.33% for Nemotron 3 Ultra. Partner figures such as Scale AI's 40x cost cut are reported by the partners themselves.

IBM's Granite Switch 4.1 comes in 8B and 30B sizes. It bundles a Decision adapter, a Guardian safety adapter and a speech model behind one endpoint, with the adapters reusing the base model's KV cache. IBM says the combined model needs about the same VRAM as the base alone.

## DeepMind and antimicrobial peptides

DeepMind describes an agent built on Google Antigravity that narrowed UniProt to a 97 amino acid bacteriocin candidate. It ran a BLAST novelty check, compared the AlphaFold structure with known bacteriocins, and drafted a wet-lab plan. The lead is in validation and no activity results are reported yet.

## What I'd watch next

Muse Spark Open benchmarks and license terms, independent Elo checks on Klein, and the Timestral technical report. For DeepMind, the first wet-lab numbers.

## My take

The pattern I keep noticing is that open weights now arrive with a serving story attached. Meta, NVIDIA and Black Forest Labs all list day-one support in the runtimes people already use. The DeepMind post is the one to read slowly, because it reports a lead and a plan but no lab result yet.

## Sources

- https://huggingface.co/blog/meta-muse-spark-open
- https://bfl.ai/blog/flux-3-image-klein
- https://mistral.ai/news/timestral
- https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning
- https://huggingface.co/blog/ibm-granite/granite-switch-4-1
- https://deepmind.google/blog/accelerating-antimicrobial-discovery

Canonical: https://heyaliasif.com/ai/2026-10-15-muse-spark-open-flux-klein-timestral