Four open-weight releases in one day: Muse Spark Open, Nemotron 3.5, Flux 3 Klein and Timestral

On 15 October 2026 four companies released models whose weights anyone can download: Meta's Muse Spark Open, NVIDIA's Nemotron 3.5 Super VL and Bolt, Black Forest Labs' FLUX 3 Image Klein and Mistral's Timestral. If the term is new, our explainer on open-weight models covers what you get and what you don't.

Most of these are mixture-of-experts models: a large pool of parameters, of which only a small "active" share does the work for each word. Memory needs follow the total, speed follows the active share. All benchmark scores below are the makers' own and have not been checked independently.

Meta: Muse Spark Open

Meta Superintelligence Labs released Muse Spark Open with 816B total and 42B active parameters and a context window of 1 million tokens. It takes text, images, video and audio as input and writes text. Meta built it for coding and agentic work. TechCrunch reported that Meta calls it the most capable open-weight model in the United States.

The hardware bar is high. The model card says it needs at least 700GB of GPU memory in the compressed NVFP4 format and 2TB in full BF16 precision. Meta also says it runs on two Mac Studios with Apple's M5 Ultra chip.

In Meta's own table it leads the open models listed on MCP-Atlas, a tool-use test, at 89.2. It trails on others: on Terminal-Bench 2.1 it scores 76.4 against 82.0 to 85.0 for Qwen3.8, GLM 5.3 and Kimi K3, and on the Artificial Analysis Intelligence Index it scores 39 against 44 and 45 for Kimi K3 and GLM 5.3. Meta stresses safety: on the StrongREJECT jailbreak test it scores 0.5, where lower is better, against 62.9 to 86.4 for the other models it lists.

The Muse Model License allows commercial use with conditions. Products must display "Built with Muse Spark Open" in the product or its documentation, and a company offering the model as a paid API service with more than US$20 million in yearly revenue must first sign a separate agreement with Meta. Downloads sit behind a sign-up form.

NVIDIA: Nemotron 3.5 Super VL and Bolt

NVIDIA released two models the same day. Super VL has 120B total and 12B active parameters, a context window of up to 1 million tokens, and reads text, images and video. Its model card lists four H100 GPUs or two GB200s for the full-precision version. NVIDIA says it is five times faster than the next-best open model in its class. Its own table is mixed: 71.1 on SWE-bench Verified, a coding test, against 79.1 for GLM-5.3-Flash.

Bolt 74B-A9B 24GB is a compressed version of Super VL with 74B total and 9B active parameters, squeezed into 24GB by storing most weights at 2 bits. It is aimed at local agents on NVIDIA RTX and DGX Spark machines, and NVIDIA lists a GeForce RTX 5090 among its test hardware. NVIDIA reports 50.98 on Terminal-Bench 2.1 against 44.38 for Qwen3.6-35B-A3B. The card also shows why self-reported numbers need care: NVIDIA measured that Qwen model at 51.2 on MCP-Atlas where Qwen reported 62.8.

Both use the OpenMDW-1.1 licence and are marked ready for commercial use. NVIDIA says a smaller Bolt E4B model follows on 23 October.

Black Forest Labs: FLUX 3 Image Klein

FLUX 3 Image Klein is an image model with 5.6B parameters in its image generator, plus a separate text encoder. It creates images from text, edits them, and combines up to 10 reference images. In BFL's internal tests it ranks first among open models and gains 91 Elo rating points in text-to-image over the older FLUX.2 [klein] 9B.

BFL says a compressed INT8 build runs on GeForce RTX cards with 8GB of memory, and makes a 1-megapixel image in about 4 seconds on an RTX 5090. The model card adds a catch: the model expects long scene descriptions, so BFL recommends a separate 9B "prompt upsampler" that expands short requests, which adds about 28GB of memory in its reference setup.

The FLUX Kommunity License allows free commercial use only for individuals and companies with under US$5 million in yearly revenue. Larger firms need a paid licence or the BFL API.

Mistral: Timestral

Timestral is not a chatbot. It is Mistral's first family of time-series forecasting models, in four sizes from 5M to 209M parameters. It predicts future values, such as sensor readings, from up to 32,768 past data points. Mistral says the 209M model ranks first on the three main public forecasting benchmarks. It needs no GPU: Mistral reports 140 milliseconds per forecast for the 209M model on an 8-thread server CPU. The licence is a "Modified MIT" licence.

Side by side

Our take

The headline model is Muse Spark Open, but for most people it is the least usable: few have 700GB of GPU memory. Bolt and Flux 3 Klein are the releases built for a single desktop GPU.

What to check before you download

1. Read the licence. Meta and BFL attach conditions tied to revenue or branding, and Mistral's is a modified, not standard, MIT licence.
2. Count the whole pipeline. Flux 3 Klein's helper model can more than double the memory it needs.
3. Wait for outside tests. Every score in this story is the maker's own; treat them as claims until others repeat them.