AI · Oct 15, 2026
Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M ContextStepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context
Oct 16, 2026, 11:29 PM · MarkTechPost
StepFun drops Step 5 Preview as open weights — 600B MoE, 27B active, 1M context — another sparse long-horizon agent model on Hugging Face.
Why it matters
MarkTechPost reports StepFun released Step 5 Preview as open weights, with FP8 and BF16 checkpoints live on Hugging Face and ModelScope. The model is described as a sparse mixture-of-experts built for long-horizon agentic work: 600 billion total parameters, 27 billion active per token, and up to 1 million tokens of context.
From the desk
Same week as Meta’s Muse Spark Open and NVIDIA’s Nemotron Super VL, StepFun is adding another open sparse giant aimed at agents. The recipe is familiar: huge total params, modest active compute, million-token context.
That’s good for teams that want to own weights for long agent sessions. It’s also more pressure on closed APIs for coding and tool-use workloads. Without independent benches in the MarkTechPost wire we have, we’re cautious on capability claims. I’m watching download traction and early SWE-style evals from the community.
Context
Open FP8 and BF16 checkpoints lower the barrier to self-hosting relative to a single proprietary format.
Who feels it
- Open-source builders
- Another long-context MoE to serve and fine-tune.
- Infra providers
- Demand for MoE-friendly inference on Hugging Face and ModelScope.
- Closed labs
- More open competition on agentic long-horizon tasks.
What to watch
- Community benchmarks against Muse Spark Open and Beam
- Serving recipes for the FP8 versus BF16 checkpoints
- Whether StepFun follows with smaller edge variants