Three Chinese laboratories dominate the open model ecosystem that rivals Claude and GPT. According to techsy.io, Qwen (Alibaba), DeepSeek and GLM (Zhipu/Z.ai) are not versions of the same model, but distinct companies with opposing technical bets.
Qwen3.6 is the broadest family, with Apache 2.0 license and lightweight footprint for self-hosting. The dense 27B parameter model runs on about 18 GB of VRAM, fitting in a 24 GB GPU. Qwen3-Coder-Next (80B-A3B) achieves 70.6-71.3% on SWE-bench Verified (self-reported).
DeepSeek V4 bets on massive Mixture-of-Experts: V4 Pro has 1.6T total parameters with 49B active; V4 Flash, 284B total with 13B active. The highlight is the price per token: V4 Flash costs US$ 0.14 per million input tokens (cache miss) and only US$ 0.0028 with cache hit. On SWE-bench Verified, V4 Pro scores ~80.6% (unverified scaffold).
GLM-5.2 is the specialist in coding and long-horizon agents. With 753B total parameters and ~40B active, MIT license and 1M context, it was tested for three weeks in production by techsy.io as the main coding model, at US$ 72/month. On Terminal-Bench 2.1 it achieved 81.0 (third party) and on SWE-bench Pro, 62.1.
In benchmarks, no model wins all categories. GLM-5.2 leads in long-horizon agent tasks; DeepSeek V4 Pro dominates aggregated coding scores; Qwen3.6-27B achieves 77.2% on SWE-bench Verified (blog report, treat with caution). In reasoning, GLM-5.2 reaches 99.2 on AIME 2025, nearly saturating the test.
Techsy.io warns that half of the flashy numbers are self-reported. A scaffold change can shift the score by a few points. The practical recommendation: GLM-5.2 for agentic coding; DeepSeek V4 Flash for cost-effective scaling with RAG; Qwen3 for local self-hosting and permissive license.
Attention: old DeepSeek endpoints (deepseek-chat, deepseek-reasoner) will be deprecated on July 24, 2026. Migrate to deepseek-v4-pro or deepseek-v4-flash.