DeepSeek V4 Flash, MiniMax H3, ByteDance Seedance 2.5: China Shifts into Third Gear in AI

Three models, a flurry of announcements over a few days at the end of July 2026, and a signal impossible to ignore: the Chinese AI ecosystem has just reached a new level. After Kimi K3 (Moonshot AI), after the continuous progress of Qwen (Alibaba), here comes the next wave: DeepSeek V4 Flash, MiniMax H3, and ByteDance Seedance 2.5.

On paper, three different products (LLM, video multimodal, video generator). In practice, the same movement: catch up and then surpass Western models in terms of performance/price ratio, systematically relying on open weights to maximize dissemination.

DeepSeek V4 Flash: the ultra-efficient LLM

Announced in preview on April 24, 2026, the DeepSeek V4 family includes two variants:

• V4 Pro: 1.6 trillion parameters in total, 49 billion active per token (Mixture-of-Experts architecture)
• V4 Flash: 284 billion parameters in total, 13 billion active per token

Both offer a context window of 1 million tokens. Both are released in open weights under the MIT license. In other words: any lab, any company, any government agency in the world can download, modify, redistribute, and host them themselves.

DeepSeek claims particularly strong performance on agentic tasks, knowledge manipulation, and inference. V4 Flash has garnered the most attention: with 13 billion active parameters, it runs on standard hardware, with inference costs reduced by an order of magnitude compared to closed models of equivalent quality.

MiniMax H3: the multimodal targeting Sora on its turf

Launched on July 30, 2026, MiniMax H3 is MiniMax's new multimodal model, with a particular focus on video generation. MiniMax claims videos that are more natural and coherent than the previous version — in AI video language, this means:

• Persistence of faces and objects across shots
• Plausible physics (gravity, fabrics, hair)
• Temporal coherence (no character "morphing" mid-shot)

This is precisely where Sora 2 (OpenAI) and Veo 3 (Google DeepMind) had taken the lead in early 2026. MiniMax is attacking this lead head-on, with the advantage of pricing.

ByteDance Seedance 2.5: the price war

On July 31, 2026 — the day after MiniMax H3 — ByteDance (the parent company of TikTok) released Seedance 2.5, its major update to the video generator. The strategic message:

• Open weights (downloadable), whereas Sora and Veo remain locked
• Improved editing functions (shot retouching, sequence coherence)
• Prices up to 7× cheaper than OpenAI or Google equivalents for professional use

The message accompanying Seedance 2.5 is clear: ByteDance doesn't want to sell a premium video model — it wants to disrupt the price market and become the default reference for creators, independent studios, and communication agencies worldwide.

MiniMax H3 on Thursday. Seedance 2.5 on Friday. In 48 hours, two leading Chinese video models, both positioned against Sora and Veo. This is no coincidence — it's tempo.

Why it matters: the reminder

With each new wave, it bears repeating: China is no longer in the race — it is the race, sometimes leading in specific segments. Four data points to measure it:

1. On open-weight LLMs, China dominates significantly

The global top 5 of open models (downloadable weights, permissive license) today is: Kimi K3 (Moonshot), DeepSeek V4, Qwen 3.5 (Alibaba), Llama 4.1 (Meta), Mistral Large 3. Three Chinese out of five, including the top two. Meta and Mistral are the only Western players keeping pace.

2. On AI video, the catch-up is swift

A year ago, Sora and Veo seemed out of reach. Today, MiniMax H3 and Seedance 2.5 are closing in on visible quality and surpass them in price/performance ratio. In content-marketing and social media uses, international creators are massively switching to Chinese tools.

3. On openness, the asymmetry is growing

Major American players (OpenAI, Anthropic, Google DeepMind) close their models. Major Chinese players open theirs. This strategic asymmetry — sometimes interpreted as a technical handicap for China (“they open because they have nothing better to sell”) — is turning around: open-weight ecosystems attract developers, researchers, startups, and universities worldwide.

4. On prices, China is driving the entire market down

ByteDance's strategy on Seedance (7× cheaper than Sora) will force OpenAI, Runway, Pika, Google to revise their pricing. Billions of dollars in margins will evaporate over the next 12 months — potentially welcome for end users, potentially problematic for Western labs' business models that rely on these revenues to fund fundamental research.

The geopolitical context at play

The US export restrictions on high-end GPU chips to China (Nvidia H100, H200, B200) are supposed to slow Chinese progress. In practice, they push Chinese players to:

• Radically optimize their architectures (MoE, quantization, distillation)
• Develop their own hardware ecosystem (Huawei Ascend, SMIC, Biren)
• Bet on openness to offset more difficult access to Western markets

The result, paradoxically: Chinese models are often more efficient per active parameter than their American counterparts. Constraint has bred ingenuity.

Key takeaways

1. The AI world is no longer bipolar (US vs Rest). It is already tripolar (US, China, Europe far behind) and is gradually shifting to bipolar (US vs China), with Europe at risk of falling out if it doesn't accelerate.
2. The three July announcements (DeepSeek V4 Flash, MiniMax H3, Seedance 2.5) should be read together: they cover LLM, multimodal, and video — three pillars of the market.
3. The cost/performance ratio is becoming the strategic axis. When an open Chinese model does 80% of what a closed American model does for 10% of the price, the market shifts.

The next episode of this quiet war is already announced: Alibaba is set to release Qwen 4 in the fall. Baidu is preparing Ernie 5. And Huawei is whispering about a “foundation” model entirely trained on Ascend hardware, without a single Nvidia chip. If the latter succeeds, the AI world map will need to be completely redrawn.