The latest Latent Space AI news roundup, which pilots a new show called High Taste Testers, leads with two US open-weights releases. Meta's Muse Spark Open is an 816B-total, 42B-active mixture-of-experts model with 1M context; Artificial Analysis gives it a 39 Intelligence Index, the highest for a US open-weights model. Reflection's Beam 1.0, released under Apache 2.0, is a 501B-total, 23B-active MoE scoring 30, placing second among US open models.
The roundup notes that Muse Spark accepts text, image, video, and audio inputs while outputting text, with BF16 weights around 1.6TB and a quantized checkpoint around 530GB. Together, Fireworks, and Baseten price it at $1.25/$4.25 per million input/output tokens. Still, the roundup flags a weak spot: Muse ranks last among the top nine open models on Artificial Analysis's four agentic evals, even as it leads open models on AA-Omniscience.
Separately, the roundup covers Gemini 4 Argon's wider availability to Pro subscribers and in AI Studio, with output limits raised to 1M tokens and intro pricing of $2/$10 until Nov 30. It also reports that Anthropic's Haiku 5.5 is about 90% cheaper than Haiku 4.5 for prompts up to 100K tokens, plus a reported Opus 5.5 fast mode and latency cuts in Claude Code. Since only this one source was provided, all claims here come from that roundup rather than independent reporting.