Observed Signal · Jul 8, 2026 · Technical Release · Source: t3n · Impact: 3/5 · Sentiment: Positive
China AI Models to Watch: Deepseek, GLM‑5.2, Qwen3.7
The article reviews recent Chinese large AI models and compares them with leading Western models, focusing on architecture, context windows, benchmark performance and token pricing. It profiles Deepseek V4 (previewed April 2026) with MoE variants (V4 Pro 1.6T, V4 Flash 248B), a Hybrid Attention Architecture and a claimed 1,000,000‑token context. Zhipu AI's GLM‑5.2 is presented as an open, agentic model suited to long‑horizon tasks with a 1M token context. Alibaba Cloud's Qwen3.7 (May 2026) ships in Max (agentic) and Plus (multimodal) editions and is now API‑only. Baidu's Ernie 5.0 (Feb 2026) is a multimodal MoE model, while Moonshot AI's open Kimi K2.6 is a high‑parameter MoE with an "Agent Swarm" design. The piece highlights competitive benchmarks and generally lower per‑token pricing versus Western counterparts.
Multiple new and updated Chinese foundation models with very large context windows, MoE architectures, competitive benchmark results and lower running costs could shift competitive dynamics, cost structures and supply of open models in the global AI/LLM market.
Track DeepSeek Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Deepseek V4 (previewed April 2026) offers MoE variants: V4 Pro (1.6 trillion parameters) and V4 Flash (248 billion); it uses a "Hybrid Attention Architecture," claims a 1,000,000‑token context, and pricing cited at $1.74 per million input tokens / $3.48 per million output tokens.
- GLM‑5.2 (Zhipu AI) is an open model positioned for long‑horizon, agentic tasks, supports a 1,000,000‑token context window, and cited operating costs of about $1.40 per million input tokens / $4.40 per million output tokens.
- Qwen3.7 (Alibaba Cloud, May 2026) is available in Max (agentic) and Plus (multimodal) variants, is no longer open‑source (access via Yotta AI Gateway or Alibaba Cloud Model Studio), with pricing for Qwen3.7 Max cited at $1.25 per million input tokens / $3.75 per million output tokens.
- Baidu Ernie 5.0 (Feb 2026) is a multimodal Mixture‑of‑Experts model (reported in the article at ~2.5 billion parameters) and has pricing cited at $1.40 per million input tokens / $5.60 per million output tokens.
- Moonshot AI's open‑source Kimi K2.6 is a Mixture‑of‑Experts model reported at ~1.04 trillion parameters with an "Agent Swarm" design; pricing noted at $0.95 per million input tokens / $4.00 per million output tokens.
Connected Companies & Entities
11 Entities mapped“Mittlerweile hat das chinesische Unternehmen Deepseek weitere Iterationen seiner KI-Modelle veröffentlicht....”
“Qwen ist das KI-Modell des chinesischen Unternehmens Alibaba Cloud....”
“Im Februar 2026 hat das chinesische Unternehmen Baidu das KI-Modell Ernie 5.0 veröffentlicht....”
“Kimi K2.6 ist ein quelloffenes Modell der chinesischen Firma Moonshot AI....”
“Schon im vergangenen Jahr äußerten sich Branchengrößen wie Nvidias CEO Jensen Huang dazu, dass China im KI-Rennen nur noch „Nanosekunden hin...”
“GLM-5.2 auf dem Niveau von Anthropics Mythos agieren soll....”
“Zum Vergleich: Für Gemini 3.1 Pro von Google werden zwei Dollar für dieselbe Anzahl an Input- und zwölf Dollar für Output-Token fällig....”
“Hier findest du externe Inhalte von TargetVideo GmbH, die unser redaktionelles Angebot auf t3n.de ergänzen....”
“Hier findest du externe Inhalte von X Corp., die unser redaktionelles Angebot auf t3n.de ergänzen....”
“Hier findest du externe Inhalte von Podigee GmbH, die unser redaktionelles Angebot auf t3n.de ergänzen....”
“Empfohlene redaktionelle Inhalte: Hier findest du externe Inhalte von YouTube Video, die unser redaktionelles Angebot auf t3n.de ergänzen....”
Ontology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Weeklong Comparison of Chinese AI Models
An indie developer spent weeks evaluating four Chinese model families—DeepSeek, Qwen, Kimi, and GLM—via Global API’s unified, OpenAI-compatible endpoint (all claim 128K context windows). The author compared pricing, latency, multimodal features and specialty strengths, then routed tasks across models to balance cost and capability for a bootcamp capstone chatbot. DeepSeek V4 Flash served as a low-cost daily default with strong code generation and ~60 tokens/sec. Qwen (Alibaba) offers broad multimodal options and very low-cost small models. Kimi (Moonshot AI) excels at multi-step reasoning and Chinese-quality outputs but is pricier. GLM (Zhipu AI) showed best Chinese-language nuance, has inexpensive small models and a vision variant. This mixed-model routing reduced monthly API spend from $400+ to about $35.
DeepSeek previews V4 open-source LLM
Deepseek on April 24, 2026 published its long‑anticipated Deepseek V4 (variants Pro and Flash), an open‑source large language model built on a new architecture with 1.6 trillion parameters. The company highlights significant gains in reasoning and autonomous code generation, claims benchmark-leading performance in mathematics, STEM and programming among open models, and says V4 supports context windows up to one million tokens while reducing compute and memory costs. Deepseek positions V4 Pro as materially cheaper on coding tasks versus OpenAI’s GPT‑5.5. The rollout also involves a partnership with Huawei, which supplies "Supernode" clusters of Ascend‑950 chips; Deepseek and analysts note a strategic focus on Huawei and Cambricon domestic chips to relieve reliance on Nvidia/AMD. Market reaction is expected to be more muted than Deepseek’s earlier 2025 breakthrough R1 shock.
China's AI Surge: Innovations Amid Controversies and Concerns
This week Chinese technology companies unveiled several new AI models spanning robotics, video generation and large language models. Alibaba’s DAMO Academy introduced RynnBrain, a robot-focused model with built-in time-and-space awareness demonstrated on tasks like object identification and multi-step manipulation. ByteDance released Seedance 2.0, a text-and-media-to-video generator praised for controllability and production quality but which suspended a feature that produced a person's voice from a photo after consent concerns. Kuaishou rolled out Kling 3.0, a subscriber-access video model claiming extended duration (up to 15s) and native multilingual audio. Separately, Zhipu AI (Knowledge Atlas Technology) published GLM-5, and MiniMax updated its open-source M2.5 model with enhanced agent tooling. Reporters note these releases position Chinese firms as closer competitors to Western video and robotics models and raise questions about content consent and model claims.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.