Skip to content
Source · Daily Brief

AI Daily Brief — 16 February 2026

A big Monday across two hemispheres. Alibaba launches Qwen 3.5 hours before Lunar New Year (Year of the Horse begins tomorrow) — a 397B-parameter / 17B-active sparse-MoE under Apache 2.0, natively multimodal across text / image / video and supporting 201 languages. Qwen 3.5-Plus on Alibaba Model Studio prices at 0.8 yuan (~$0.12) per million tokens — roughly 1/18 the cost of Google’s Gemini 3 Pro. In New Delhi, PM Modi inaugurates the India AI Impact Expo 2026 at Bharat Mandapam, opening the India AI Impact Summit (16-21 February) — the first such summit hosted by a Global South nation.

Top stories

  • Alibaba — Qwen 3.5 ships open-weight. 397B total / 17B active MoE. Native multimodal (text + images + video in one model), 201 languages and dialects (up from 82 in Qwen 3, now including Hawaiian and Fijian). Hybrid architecture fuses Gated Delta Networks (linear attention) with sparse MoE for an 8.6× → 19× decoding speed-up vs Qwen3-Max at ~60% lower cost. Weights on Hugging Face and ModelScope under Apache 2.0. via Qwen
  • Qwen 3.5-Plus undercuts Gemini 3 Pro ~18×. Available on Alibaba’s Model Studio with API pricing at 0.8 yuan (~$0.12) per million tokens. A more powerful Qwen 3.5-Max variant and additional family members are telegraphed to follow. via Caixin Global
  • Modi opens India AI Impact Expo 2026. The curtain-raiser for the India AI Impact Summit (16-21 February) at Bharat Mandapam. The Expo runs through 20 February across 70,000+ sq m with 300+ exhibition pavilions, 13 country pavilions (Australia, Japan, UK, France, Germany, Russia, a collective Africa pavilion). Structured around three “Sutras” — People, Planet, Progress. via PMO India
  • India unveils whole-of-nation AI strategy. IT Minister Ashwini Vaishnaw outlines a “frugal, sovereign and scalable” AI ecosystem covering the five layers of the AI stack (applications, models, compute, talent, energy). The government commits to add more than 20,000 GPUs on top of the existing 38,000 under the IndiaAI Compute Portal; AI Mission 2.0 is framed as the “UPI of AI” for MSMEs, healthcare and education. via Wikipedia / DD News

Open-source pulse

Qwen 3.5’s launch caps an extraordinary China release week ahead of Lunar New Year: ByteDance’s Doubao-Seed-2.0 + Seedance 2.0 on 14 February (positioned vs GPT-5.2 / Gemini 3 Pro), Zhipu’s GLM-5 on 11 February (744B / 44B active MoE, MIT-licensed, Huawei Ascend-trained), with DeepSeek V4 still rumoured around the holiday. The China stack now occupies multiple top spots on Artificial Analysis open-weights ladders. via Euronews

Policy & markets

The Bank of England publishes a summary of its AI roundtables with banks and insurers: industry broadly backs the PRA’s principles-based approach (Supervisory Statement 1/23 on Model Risk Management) but flags that traditional MRM and validation may not scale to generative / agentic AI, and questions how “human-in-the-loop” applies as AI takes on more decision-making.

Quiet corners

Despite Anthropic’s Thursday $30B / $380B close, no new Western frontier-lab model launches on Monday. OpenAI, Anthropic, Google DeepMind, Meta AI are quiet on first-party shipping; the 17-19 February cluster (Anthropic Claude Sonnet 4.6, xAI Grok 4.20, Google Gemini 3.1 Pro, Meta × NVIDIA) is queued. The day belongs to China and India.

By the numbers

  • Qwen 3.5: 397B / 17B active MoE · 201 languages · Apache 2.0 · 8.6-19× decoding speed-up vs Qwen3-Max
  • Qwen 3.5-Plus API pricing: $0.12 / M tokens (~1/18 of Gemini 3 Pro)
  • India AI Impact Expo: 70,000+ sq m · 300+ pavilions · 13 country pavilions · 5 days
  • IndiaAI Compute Portal: +20,000 GPUs on top of existing 38,000

Compiled by AI Feed’s editor from publicly reported announcements on 16 February 2026.