Domestic vs. International Mainstream AI Models: A Full-Spectrum Comparison of Text, Multimodality, and Commercialization (2026)

By 2026, the global large-model race has extended far beyond pure text generation into the deep waters of "full-modal understanding and generation." On the domestic front, Kimi (Moonshot AI), MiniMax, Zhipu AI, and Alibaba (Tongyi Qianwen) are going head‑to‑head with OpenAI, Anthropic, and Google in the text arena. In video generation, ByteDance (Seedance) and MiniMax (H3) are flying the flag for China, competing fiercely with international rivals like Sora and Veo. This article provides a systematic comparison of the latest domestic and international models from two key angles: model capabilities and token pricing.

I. Text Model Capabilities: Domestic Breakthroughs in Niche Areas, but Global Leaders Still Ahead Overall

1. Four Major Domestic Text Players and Their Latest Flagships

(1) Kimi (Moonshot AI): K3 – Largest Parameter Count, World‑leading Coding Performance

In July 2026, Moonshot AI released Kimi K3, a model with 2.8 trillion parameters – the first open‑source 3‑trillion‑level model globally. Built on a self‑developed KDA hybrid linear attention mechanism, it achieves roughly 2.5× higher scaling efficiency than its predecessor. On the Arena AI Frontend Code Arena, K3 tops the leaderboard with 1679 points, surpassing Claude Fable 5 (1631) and GPT‑5.6 Sol (1618). In Vals AI’s comprehensive tests, it scores 74.7, ranking second worldwide. Its main drawbacks are relatively slow inference speed and the fact that full capabilities require the highest‑tier subscription. It supports a 1‑million‑token context window.

(2) MiniMax: M3 – The Open‑Source Differentiator with "Triple‑Threat" Capabilities

In June 2026, MiniMax unveiled M3, positioned as China’s first open‑source model that simultaneously delivers "cutting‑edge coding, 1M long context, and native multimodality." With 428 billion total parameters and about 23 billion activated per token, its self‑developed MSA (MiniMax Sparse Attention) architecture accelerates prefilling by over 9× and decoding by over 15× under a 1M context. On SWE‑Bench Pro, M3 surpasses GPT‑5.5 and Gemini 3.1 Pro. MiniMax’s total revenue in 2025 reached $79.038 million, with overseas revenue accounting for 73%.

(3) Zhipu AI: GLM‑5.2 and GLM‑5.3 – Dual Breakthroughs in Long‑Horizon Tasks and Security

• GLM‑5.2 (released in June): Focuses on long‑horizon tasks, enabling AI to autonomously complete an entire engineering workflow from development and integration to testing and deployment. With a 1M context, it trails Claude Opus 4.8 by only about 1‑4% on benchmarks like FrontierSWE. Notably, on Day 0 it completed inference adaptation for domestic computing platforms including Huawei Ascend, Pingtouge, and Moore Threads.

• GLM‑5.3 (released in August): Further improves intelligence through post‑training, achieving open‑source No. 1 on Terminal Bench 3.0, Agents’ Last Exam, and others. It scores 84.5% on the CyberGym vulnerability reasoning benchmark (vs. 77.2% for GLM‑5.2), and with its help researchers uncovered a DNS‑level vulnerability that had existed for over 40 years.

(4) Alibaba (Tongyi Qianwen): Qwen 3.8‑Max and Qwen3.8‑Flash – A Dual‑Track Strategy

• Qwen 3.8‑Max (released in August): With 2.4 trillion total parameters and a sparse MoE architecture (activating 95 billion at inference), it supports a 1M context. After a September iteration, it topped the CodeArena WebDev leaderboard with 1691 points, ahead of Claude Opus5 and Kimi K3.

• Qwen3.8‑Flash (open‑sourced in August): A 100‑billion‑parameter model that activates only 6 billion, yet outperforms Claude Opus4.6 by 9.1 points on SWE‑bench Pro. Its training cost is nearly 90% lower than Qwen3.7‑Plus.

2. The Three Major International Players: The Strong Moat of Closed‑Source Ecosystems

Interim assessment: Domestic models have achieved head‑to‑head competitiveness in specialised areas such as coding and long‑horizon tasks. However, overall comprehensive ability and user experience still lag behind Claude Fable 5 and GPT‑5.6 Sol.

II. Multimodal/Video Generation Capabilities: A Two‑Horse Race, with Domestic Models Now Among the Global Elite

By 2026, the video generation track has formed a "ByteDance Seedance vs. MiniMax H3" duopoly, with Alibaba, Zhipu, Kuaishou and others also competing vigorously.

1. ByteDance Seedance Series: The Industrial‑Grade Benchmark

• Seedance 2.0 (released in February): Topped the Artificial Analysis Video Arena with Elo 1269, surpassing Google Veo 3, OpenAI Sora 2, and Runway Gen‑4.5. Its self‑developed Dual‑Branch Diffusion Transformer (DB‑DiT) synchronously outputs visuals and audio in a single forward pass, eliminating temporal misalignment. It supports up to 60 seconds of 2K video, lip‑sync in 8+ languages, and multimodal reference inputs (9 images + 3 videos + 3 audio), with a 90% availability rate.

• Seedance 2.5 (released July 31): Four major upgrades – single generation extended to 30 seconds, support for native 4K and 10‑bit colour; multimodal reference capacity raised to up to 50 full‑modal assets, with new white‑model and green‑screen capabilities; added local video editing. On release day, it secured partnerships with Sany Group, XPeng Motors, and others.

• Seedance 2.0 mini (launched in June): A cost‑effective version at 0.023 yuan per thousand tokens for image‑to‑video, roughly 0.5 yuan per second.

2. MiniMax H3: The Open‑Source Full‑Modal Disruptor

MiniMax H3, released on July 31 and open‑sourced on August 3, is MiniMax’s first open‑source multimodal generation model. It can jointly understand text, image, video and audio contexts, and generate up to 2K resolution, 15‑second videos with native stereo audio. On Artificial Analysis, it ranks first globally in video editing, and second/third in text‑to‑video and image‑to‑video respectively. It accepts up to 9 images + 3 videos + 3 audio as multimodal references. Pricing is only one‑third of industry peers – 0.8 yuan/second for 2K, 0.5 yuan/second for 768P.

Head‑to‑head comparisons show H3 and Seedance 2.5 are "equally strong but with distinct personalities" – H3 understands better what the final video should "look like," while Seedance 2.5 knows better what the director specifically "wants to do."

3. Other Domestic Multimodal Models at a Glance

4. International Multimodal Counterparts

III. Full‑Spectrum Token Pricing Analysis: From Text to Multimodal

1. Text Model Pricing: Domestic "Roller‑Coaster" vs. International Stability

Domestic – From "Price War" to "Collective Hikes"

In the first half of 2026, domestic text models endured a brutal price war, with DeepSeek V4‑Pro’s cache‑hit price dropping to as low as 0.025 yuan per million tokens, and many mainstream models cutting prices by over 90%. But in the second half, driven by explosive Agent usage (weekly token volume peaking at 7.22 trillion), GPU shortages, and a >500% surge in HBM prices, the industry collectively raised prices sharply. According to a Morgan Stanley August report, average Chinese API input and output prices rose by about 48% and 80% year‑on‑year, respectively.

Key adjustments:

• Kimi K3: Output price raised to 100 yuan/million tokens (~$14), a >3.5× increase from K2.6 (22.4 yuan); input at 20 yuan/million (2 yuan cache‑hit).

• Zhipu GLM: Three price hikes this year, average increase 83%.

• Alibaba Qwen 3.8‑Max: Input 12 yuan, output 36 yuan per million tokens.

• Alibaba Qwen3.8‑Flash: Input 1 yuan, output 3 yuan per million tokens – only 3% of Claude Opus4.6.

International – Stable and High, with Clear Premium

Observation: Kimi K3’s output price of ~$14/M tokens is still significantly lower than Claude Fable 5’s $50 and GPT‑5.6 Sol’s $30. Domestic models retain a price advantage even after hikes, but the gap is narrowing. The hikes reflect the painful but necessary shift from "burning cash for market share" to "sustainable commercialisation."

2. Multimodal/Video Model Pricing: Per‑Second vs. Per‑Token Logic

Video pricing is fundamentally different from text – it depends on resolution, duration, and input complexity, making direct comparison difficult.

Seedance 2.5 is about 50% more expensive than 2.0, linking cost directly to content complexity – simpler scenes consume fewer tokens and are cheaper, while video‑input editing/continuation tasks align with industry per‑second rates. MiniMax H3’s 0.8 yuan/second is only one‑third of flagship peers; a 15‑second 2K video costs only about $1.95, whereas Seedance 2.5 would cost roughly 10 yuan/second – H3’s cost advantage is enormous.

IV. Core Trends and Differentiating Paths for Domestic Models

1. Paradigm Shift from "Single‑Modal" to "Full‑Modal"

By 2026, the competitive benchmark has moved from "being able to generate video" to "understanding and generating multimodal content holistically." Seedance 2.5 supports 50 full‑modal reference assets; MiniMax H3 unifies text/image/video/audio as freely callable context; Alibaba Wan3.0 is the first to accept office documents for direct video generation. "Full‑modal understanding + full‑modal generation" is becoming the new standard.

2. Open‑Source as a Key Differentiator for the Domestic Camp

While leading international models (GPT‑5.6, Claude Fable 5) remain closed, domestic players are collectively embracing open source: Kimi K3, MiniMax M3 & H3, Zhipu GLM‑5.2/5.3, and Alibaba Qwen3.8‑Flash are all open‑sourced. In video generation, MiniMax H3’s open‑source approach echoes DeepSeek’s disruptive path, and is likely to attract a global developer ecosystem.

3. Domestic Computing Adaptation: A Strategic Moat

Zhipu GLM‑5.2 was the first to complete inference adaptation for multiple domestic computing platforms including Huawei Ascend. Under the "bottleneck" of Nvidia’s high‑end GPUs, the combination of open‑source domestic models + domestic computing power is building a unique strategic advantage, ensuring supply‑chain resilience amid Sino‑US tech tensions.

4. Video Generation: From "Creative Toy" to "Productivity Tool"

Seedance 2.5 has already been deployed in real‑world industrial scenarios: Sany Group (industrial training SOP videos), XPeng Motors (design visualisation), and QiaoXuan Intelligence (embodied‑AI training data generation). Video generation is transitioning from "generating content" to "understanding the world," and from "aiding creativity" to "boosting productivity."

V. Conclusion

In 2026, domestic large models are no longer just "chasers." In the text arena, Kimi K3 and Qwen 3.8‑Max have topped global leaderboards in coding and other specialised domains; although overall capabilities still trail Claude and GPT’s top closed‑source models, the gap continues to narrow. In the multimodal arena, Seedance sets the industrial benchmark with 4K/30‑second generation and enterprise‑ready delivery, while MiniMax H3 takes a different path with open‑source and extreme cost‑performance – both have secured a place among the global first tier.

On pricing, the text market’s swing from price wars to collective hikes signals the industry’s maturation from land‑grabbing to sustainable commercialisation. Multimodal pricing shows a diversified landscape with both per‑token and per‑second models, with H3 standing out for its cost advantage.

The ultimate competition in large models has evolved from a narrow battle of model capabilities into a comprehensive contest spanning text and multimodal technology depth, open‑source ecosystem building, domestic computing adaptation, and commercial pricing strategy. Domestic players now lead in specific niches, but achieving full‑spectrum leadership will require continued innovation in foundational architecture and further refinement of user experience.