Skip to main content
Back

Z.ai's GLM-4.7-Flash Beats Qwen3 on Coding Agents at $0.06 per Million Tokens

Z.ai's 30B mixture-of-experts model activates only 3B parameters yet beats Qwen3-30B and GPT-OSS-20B on agentic coding benchmarks.

  • Z.ai released GLM-4.7-Flash, a 30B-A3B MoE model with only 3B active parameters, MIT licensed.
  • Scores 59.2% on SWE-bench Verified and 79.5% on τ²-Bench, beating Qwen3-30B and GPT-OSS-20B on agentic tasks.
  • Uses Multi-head Latent Attention to compress KV cache by up to 93%, enabling a 202K token context.
  • Three thinking modes (interleaved, preserved, turn-level) preserve reasoning across multi-turn agent sessions.
  • Runs at ~19.8 GB VRAM in Q4_K_M; served via vLLM, SGLang, DeepInfra at $0.06/$0.40 per million tokens.
  • Positioned as the free-tier Haiku-equivalent in the GLM-4.7 series, with a $10/month coding plan.

Z.ai has quietly slipped a new small model into its GLM-4.7 lineup, and the numbers are turning heads. GLM-4.7-Flash is a 30B-A3B MoE model positioned as the strongest option in the 30B class, aimed at lightweight deployment that balances performance and efficiency. In practice that means 30 billion total parameters but only around 3 billion active per token, putting it in the same weight class as Qwen3-30B and OpenAI's GPT-OSS-20B while beating both on most agentic benchmarks.

Z.ai pitches Flash as a lightweight, free-tier version of GLM-4.7 that delivers strong performance across coding, reasoning, and generative tasks with low latency and high throughput. It ships under an MIT license on Hugging Face, and inference providers like DeepInfra are already serving it at roughly $0.06 per million input tokens and $0.40 per million output tokens, aggressive pricing even by open-weight standards.

Benchmarks that don't add up on paper

What makes Flash interesting is that it dominates on the tasks people actually build agents for, even when it doesn't win on math or general knowledge. On the model card, Z.ai reports these head-to-head numbers against the closest open competitors:

BenchmarkGLM-4.7-FlashQwen3-30B-A3B-ThinkingGPT-OSS-20B
AIME 25 (math)91.685.091.7
GPQA (science)75.273.471.5
HLE (Humanity's Last Exam)14.49.810.9
SWE-bench Verified59.222.034.0
τ²-Bench (tool use)79.549.047.7
BrowseComp42.82.2928.3

The SWE-bench Verified score of 59.2% is the eye-catcher. That is nearly triple Qwen3-30B's number and comfortably above GPT-OSS-20B, on a benchmark that requires resolving real GitHub issues in real repositories. The τ²-Bench and BrowseComp gaps are even wider, suggesting Z.ai has specifically optimized this model for multi-turn tool use and web navigation over one-shot Q&A.