Skip to content
Source · Daily Brief

AI Daily Brief — 6 November 2025

Thursday compressed two structural shifts. Moonshot AI released Kimi K2 Thinking on Hugging Face — 1T total params / 32B active MoE, 256K context, native INT4 quantization, sustains 200-300 sequential tool calls, modified MIT license (attribution required >100M MAU or >$20M monthly revenue). SOTA on agentic-reasoning benchmarks: Humanity’s Last Exam (tools) 44.9% vs GPT-5’s 41.7% and Claude 4.5’s 32.0%; BrowseComp 60.2%; SWE-Bench Verified 71.3% — first open-weight model to lead these. Training cost reported at ~$4.6M one-shot (Moonshot CEO later said “not official” since it ignores research/failed experiments, but the number reset open-source efficiency expectations). API at $0.60/$2.50 per Mtoken in/out. Microsoft launched MAI Superintelligence Team under Mustafa Suleyman — explicitly “Humanist Superintelligence” (HSI) framing, anti-AGI-race posture: “subservient to humans,” problem-oriented rather than “ill-defined and ethereal.” Joined by Chief Scientist Karén Simonyan. Focus areas: digital companions, disease diagnosis, renewable energy. Analysts read it as Microsoft asserting strategic autonomy from OpenAI. Inception Labs (Stanford prof Stefano Ermon) closed $50M seed for diffusion LLMs — Menlo lead with Mayfield, Innovation Endeavors, M12, Snowflake, Databricks, NVentures; Ng + Karpathy as angels. New Mercury coding model claims 10× speedup vs autoregressive. Qualcomm + Arm Q3 results landed (Nov 5) — both beat on AI demand, mixed market reaction.

Top stories

  • Moonshot releases Kimi K2 Thinking. 1T MoE / 32B active / 256K / INT4 native / 200-300 sequential tool calls. Modified MIT. via Hugging Face
  • K2 Thinking tops GPT-5 + Claude 4.5 on agentic benchmarks. HLE w/tools 44.9% (vs GPT-5 41.7%, Claude 32.0%), BrowseComp 60.2%, SWE-Bench Verified 71.3%. First open-weight to lead these. via VentureBeat
  • K2 Thinking ~$4.6M training cost (unofficial). Orders of magnitude below US frontier-lab budgets. Moonshot CEO later said figure ignores research/failed experiments — but the number reset open-source efficiency expectations. via Simon Willison
  • K2 Thinking INT4-native at trillion-param scale. First major frontier-class open-weight release native INT4 — Moonshot claims lossless inference latency/memory wins. Combined with 256K + tool-loop budget, redefines deployable open-source upper bound. via Moonshot AI
  • Microsoft launches MAI Superintelligence Team under Suleyman. “Humanist Superintelligence” — anti-AGI-race, problem-oriented, “subservient to humans.” Karén Simonyan as Chief Scientist. Focus: digital companions, disease diagnosis, renewable energy. via CNBC
  • Suleyman: “We reject narratives about a race to AGI”. “Towards Humanist Superintelligence” blog lays out philosophical break — frames superintelligence as “subservient to humans” and problem-oriented rather than “ill-defined and ethereal.” via Microsoft AI
  • Inception Labs $50M seed for diffusion LLMs. Ermon (Stanford). Menlo lead with Mayfield, Innovation Endeavors, M12, Snowflake, Databricks, NVentures. Ng + Karpathy as angels. New Mercury model claims 10× speedup vs autoregressive. via TechCrunch
  • K2 Thinking framed as “DeepSeek moment”. Multiple outlets converge on “most powerful open-source thinking model” — first open-weight to lead HLE + BrowseComp with native thinking-while-using-tools support over hundreds of sequential calls. via AI News
  • Microsoft signals autonomy from OpenAI. Suleyman-led superintelligence org — building proprietary frontier capabilities rather than relying solely on OpenAI’s roadmap. via Axios

Who shipped

Moonshot AI shipped Kimi K2 Thinking. Microsoft shipped MAI Superintelligence Team. Inception Labs shipped $50M seed + Mercury model. OpenAI, Anthropic, Google, Meta, xAI quiet on launches.

Open-source pulse

K2 Thinking restructures open-source ceiling overnight. Combination of trillion-param scale + INT4 native + 256K + 200-300 sequential tool calls + modified MIT is unprecedented. ~$4.6M training claim, even unofficial, pressures US-lab training-cost narratives.

Money, infra & hardware

Qualcomm FY-Q4 EPS $3.00 (vs $2.88 est) on $11.27B (+10% YoY), Q1 FY26 above consensus. Arm Q2 FY26 $1.14B (+34% YoY) — third straight $1B+ quarter, above guidance high end. Inception Labs $50M is the day’s frontier round.

Quiet corners

Microsoft’s HSI framing is the substantive policy moment — first major-lab leader explicitly rejecting AGI-race framing as worth pursuing.

By the numbers

  • 1T / 32B / 256K / INT4 — K2 Thinking total / active / context / native quantization
  • 44.9% / 60.2% / 71.3% — HLE w/tools / BrowseComp / SWE-Bench Verified
  • ~$4.6M / 200-300 — training cost (unofficial) / sequential tool-call sustain
  • $0.60 / $2.50 — K2 Thinking API per Mtoken in/out
  • $50M / 10× — Inception Labs seed / claimed diffusion-LLM speedup
  • Most-mentioned company: Moonshot AI
  • Quietest segment: OpenAI launches

Compiled by AI Feed’s editor from verified web sources for 6 November 2025.