Skip to content
Source · Daily Brief

AI Daily Brief — 21 March 2025

GTC closed and China answered. Tencent launched Hunyuan-T1, the first ultra-large hybrid Mamba-Transformer MoE reasoning model — pricing it at roughly 1% of OpenAI o1, matching DeepSeek-R1 on math benchmarks, and beating GPT-4.5 on MMLU-Pro. NVIDIA wrapped its week with the open AgentIQ orchestrator. Adobe Summit drove the enterprise agentic story.

Top stories

  • Tencent releases Hunyuan-T1 — first ultra-large hybrid Mamba-Transformer MoE reasoning model. Built on TurboS base (56B activated / 560B total params), decodes at 60-80 tokens/sec (2x its Transformer rivals), 128K context. Scored 87.2 on MMLU-PRO (vs DeepSeek-R1 84.0, GPT-4.5 86.1, o1 89.3), 96.2 on MATH-500, 93.1 on LiveCodeBench logic. Pricing: 1 yuan per million input / 4 yuan per million output (~$0.14 / $0.55). via Tencent
  • Hunyuan-T1 reframes the price war. Roughly 99% cheaper than OpenAI o1 at output for comparable math/code performance, and undercuts DeepSeek-R1 on inference speed via the Mamba-Transformer hybrid. T1-Preview had shipped on Tencent’s Yuanbao app in mid-February; Mar 21 is the official release. via Analytics Vidhya
  • NVIDIA closes GTC with AgentIQ open-source agent toolkit. Framework-agnostic library for orchestrating disparate agents and frameworks (LangChain, LlamaIndex, CrewAI etc.) — treats agents/tools/workflows as composable function calls, emits granular telemetry to identify bottlenecks. Permissive licensing, no obligation to share modifications back. via InfoWorld
  • Adobe Summit launches agentic AI platform with 10 prebuilt agents. At its Las Vegas Summit, Adobe unveiled an agentic platform covering audience optimization, personalization, content production and journey orchestration. Parallel data-sharing partnerships with AWS, Microsoft and SAP widen training-data access across Adobe Experience Platform. via MarketingProfs

Who shipped

Tencent took the headline — Hunyuan-T1 is the loudest Chinese reasoning-model release since DeepSeek-R1 and shows Beijing has multiple frontier-tier shops. NVIDIA closed GTC with AgentIQ. Adobe and Cisco ran the enterprise-agent narrative.

Open-source pulse

AgentIQ shipped open-source under a permissive license. Hunyuan-T1 itself was API-released but Tencent has open-sourced earlier Hunyuan checkpoints; weights for T1 were still pending at launch.

Money, infra & hardware

NVIDIA highlighted new robotic-surgery and autonomous-imaging projects on GTC’s closing day, building on the Isaac GR00T N1 foundation model unveiled earlier in the week.

Quiet corners

A Columbia Journalism Review Tow Center study published this week found that ChatGPT Search, Perplexity, Gemini, Copilot and Grok-3 returned incorrect citations in over 60% of tested queries — sometimes fabricating URLs or misattributing copied text. More fuel for the publisher-vs-AI licensing fights.

By the numbers

  • 56B / 560B — Hunyuan-T1 activated / total parameters
  • 60-80 tok/s — decoding speed (2x Transformer rivals)
  • 87.2 / 96.2 / 93.1 — MMLU-PRO / MATH-500 / LiveCodeBench logic scores
  • ~$0.14 / $0.55 per million tokens — Hunyuan-T1 input/output pricing
  • ~99% cheaper than OpenAI o1 at comparable benchmarks
  • 10 — prebuilt agents Adobe shipped at Summit
  • >60% — error rate for AI-search citations per CJR Tow study
  • Most-mentioned company: Tencent

Compiled by AI Feed’s editor from verified web sources for 21 March 2025.