How Chinese LLMs Are Closing the Gap: DeepSeek V4, Kimi K3, and GLM Updates

This third installment of the "ALL IN AI" column surveys the latest developments of Chinese large language models (LLMs). Over the past two months DeepSeek released V4 Pro (model ID DeepSeek-V4-Pro-0813) and V4 Flash ( DeepSeek-V4-Flash-0731), Kimi launched K3 with 2.8 T parameters, native vision and a 1 M token context window, and the GLM Coding Plan became fully subscribable.

The author emphasizes that while Chinese LLMs have not yet universally outperformed top overseas models such as Anthropic’s Claude or OpenAI’s GPT, the performance gap is shrinking rapidly. Model capabilities vary across dimensions—reasoning, knowledge, code generation, vision, tool use, long context, speed, and stability—so a single metric cannot capture overall quality.

DeepSeek V4 Pro vs. V4 Flash : Both offer 1 M context, tool calls, JSON output, and OpenAI/Anthropic‑compatible APIs. V4 Pro (1.6 T total, 49 B active parameters) targets complex reasoning and high‑quality output, while V4 Flash (284 B total, 13 B active) prioritizes faster response, higher concurrency, and lower cost. Pricing (as of 2026‑08‑14) per million tokens is 0.02 ¥ (cache hit) and 1 ¥ (cache miss) for Flash, 0.025 ¥ and 3 ¥ for Pro, with output costs of 2 ¥ and 6 ¥ respectively. A peak‑off‑peak pricing scheme starts on 2026‑08‑17, doubling cache‑miss and output rates during peak hours.

Kimi K3 : The flagship model provides 2.8 T total parameters, 1 M token context, native visual understanding, and three inference intensity levels ( low, high, max). Two model IDs exist: k3 (1 M context for long‑form tasks) and k3-256k (256 K context for routine queries, consuming fewer quota units). Pricing is 0.30 USD (cache hit), 3.00 USD (cache miss), and 15.00 USD per million output tokens, making K3 more expensive than DeepSeek but offering visual capabilities.

GLM Coding Plan : Previously limited by supply, the plan now offers unrestricted subscription across three tiers (Lite, Pro, Max) with significant price increases (e.g., Pro from 149 ¥ to 538 ¥). The plan supports GLM‑5.2, GLM‑5‑Turbo, and GLM‑4.7, providing vision, web search, and code‑related MCP abilities. Pricing varies by time of day, with higher rates during peak periods (14:00‑18:00 UTC+8).

Benchmark comparison : Using the Artificial Analysis Intelligence Index v4.1.1 (nine tests including GDPval‑AA, Terminal‑Bench, SciCode, etc.), scores on 2026‑08‑14 data show Claude Fable 5 at 60, GPT‑5.6 Sol at 59, Kimi K3 (max) at 57, DeepSeek V4 Pro at 53, DeepSeek V4 Flash at 52, and GLM‑5.2 at 51. This places Chinese models within the same high‑performance band as leading foreign models, though gaps remain.

The author warns against judging a model’s quality from a single conversation or anecdotal experience. Six factors influencing outcomes are listed: task clarity, sufficient context, appropriate tools, matching inference budget, verifiable execution, and model suitability. A robust evaluation should combine transparent public benchmarks, identical real‑world tasks, and analysis of quality, speed, stability, and cost.

In conclusion, Chinese LLMs have progressed to a stage where they offer competitive capabilities, lower pricing, and longer context windows, narrowing the gap with overseas leaders. Users should select models based on their specific workloads, budget, and required features rather than chasing a single "best" model.