Fish Audio & Meta Dominate: The Dual Narrative on X
Open-weight audio stole the week on X. Fish Audio's Drama 3 launch, a 35B-parameter text-to-speech model with open weights and 83 languages, pulled 6,885 posts and an 8.6M-view launch post, with the Uncle Roger clone demo doing most of the carrying. Meta's Muse Wonder had the single most-liked post of the window (38.4K likes, 5.4M views on Zuckerberg's announcement), but the thread that mattered for builders was AI at Meta clarifying that Wonder stays closed while Muse Spark weights are "coming soon." One more thing…. Free to read and Pay when it acts. 30% extra on every top-up through Oct 31 2026. Claim 30% Bonus.[Sponsored]
Below the 2 giants, the chatter was all about local and agentic: Liquid AI's LFM2.5-24B-A2B (24B MoE, 2B active, 131K context) at 1,408 posts, DeepMind's Gemini 4 Argon teaching an Apptronik humanoid cornhole from 1 video at 1,357 posts, and Exa crossing $100M ARR on 2B agent search requests a month at 684 posts. The 1 non-model story worth watching is hardware: China's CXMT 4F² DRAM news knocked Micron, Samsung and SK Hynix and had the most-viewed finance post in AI circles at 628K views.
Featured
▲ Google DeepMind released the Gemini 4 Argon API with a 1M-token output limit. 2 weeks ago Argon was Fairwind Program only. Now it is GA for developers, enterprises, and paid Gemini subscribers. It is a sparse MoE built for long multi-step work: coding, finance, legal, science, cyber. 1 response can run to 1M tokens, up from 64K. Input and output share 1 1M window, so a long prompt leaves less room to write. Pricing is $2 input, $10 output per 1M tokens through November 30, then $4/$20, which matches Claude Opus 5.5. Vals AI ranks it first on the Vals Index at 68.90%. LVBench is 91.7%. It trails GPT-6 Astra on FrontierSWE v2, 55.0% vs 65.5%, and Opus 5.5 on Terminal-Bench 4.0, 57.4% vs 66.4%. Argon loses 5 rows of Google's own table. Rival scores are self-reported or from public boards. Google's DeepSWE run used its own harness.
Also this week’s main signals
▲ Meta Launches Muse Wonder and Muse Wonder Full: Benchmarks, Pricing and How to Use It. Meta's most capable models so far, per Meta. Both take text, image, video, and audio with 1M context. Muse Wonder is $4 input, $20 output per 1M tokens on the Meta Model API and Muse Code. Muse Wonder Full is paid Muse Code only: $50/month Power or $200/month Maximum. On DeepSWE 1.1, Wonder Full scores 79.0 to Opus 5.5's 74.2. Wonder scores 85.6 on MemoryBench, a Meta-internal eval.
▲ 🐟 🐟 Free to read and Pay when it acts. 30% extra on every top-up through Oct 31 2026. Claim 30% Bonus [Sponsored]
▲ Liquid AI Releases LFM2.5-24B-A2B: A 2B-Active MoE for On-Device Agents 24B total, about 2B active per token, 131K context, text only, reasons before every answer. Liquid trained it inside agent harnesses for multi-step tool use. On AutomationBench it scores 27.69 to Qwen3.6-35B-A3B's 8.72. On SWE-bench Verified it trails, 51.00 to 66.20. Liquid's own runs on vLLM.
▲ Datalab released OmniExtractBench, an open benchmark for structured extraction from PDFs. It pools 620 documents from 4 sources, so no single vendor picked the set. 33 documents run past 100 pages.
Other Signals
▲ Anthropic's Frontier Red Team ran a 26B model across 15 Android TV boxes at $55 each, 0.39 tokens/s.
▲ AMD's Ryzen AI Halo Developer Platform has 192GB unified memory, 160GB allocatable as VRAM, ordering now.
▲ 🐟 🐟 Free to read and Pay when it acts. 30% extra on every top-up through Oct 31 2026. Claim 30% Bonus [Sponsored]
▲ NVIDIA contributed Vera Rubin rack designs to OCP via MGX 1.2: 5,000A busbar, racks to 250kW.
▲ Meta's Muse is now on iPad with Asana, Figma, GitHub, and Zoom connectors, weeks after topping iPhone.
▲ AMD Helion is an open agentic-AI rack blueprint, up to 32,768 EPYC cores, partner systems due November.
▲ Amex's Intent Intelligence Framework issues a Proof of Intent Token, so agents never touch your card number.
Until next week,
The Marktechpost Team
Sponsorship Opportunities