Model details
GLM-4.7 Flash takes a practical dense-architecture approach rather than the massive MoE design of its flagship sibling. With its 31-billion-parameter dense structure, the model activates all parameters for every token generated, producing highly predictable inference behavior without the latency spikes that can plague larger sparse systems. This architectural choice makes the model far easier to quantize and deploy locally, and it enables the model to hit 80+ tokens per second on consumer hardware like MacBooks—a meaningful figure for developers who want meaningful AI assistance without renting time from massive proprietary systems. The design philosophy here is pragmatic workhorse rather than all-knowing AGI: the model is explicitly built for teams and individuals who need reliable code generation and reasoning but are constrained by budget or hardware.
The model shows its strongest identity in coding and agentic scenarios, achieving 59% on Software Engineering Bench, 79.5% on TA2 agentic tasks, and 75.2% on GPQA—scores that position it as a credible alternative where proprietary models would otherwise dominate budgets. Released under an MIT license with free API access, GLM-4.7 Flash is designed for the full development lifecycle: from experimenting with ideas to shipping real tools, making agentic workflows viable for smaller teams that could never justify the cost of massive frontier models. The 128,000-token context window allows developers to feed entire repositories, heavy documentation, or lengthy stack traces into a single conversation, and the dense architecture simplifies local quantization for teams that want to run their own infrastructure. This combination of open weights, coding-focused benchmarks, and hardware-friendly efficiency makes the model a natural fit for developers who want frontier-level coding ability without frontier-level costs.
Transparent token rates
Rates are shown per one million tokens. Combined means one million input plus one million output tokens.
Amazon Bedrock
Z.AI's GLM-4.7 Flash is a 31-billion-parameter open-source model released for coding, reasoning, and agentic workflows, offering free API access and local deployment. It records 59% on Software Engineering Bench, 79.5% on TA2 agentic tasks, and 75.2% on GPQA, while pricing lists $0.07 input, $0.01 cached input, and $0.
Amazon Bedrock
GLM-4.7 Flash packs 31B parameters and an MIT license with free API access, helping you test ideas and ship tools on a tiny budget.
Amazon Bedrock
The 30B model achieves 59.2% on SWE-bench Verified while running at 80+ tokens per second on MacBooks.
Amazon Bedrock
The news blog specialized in Japanese culture, odd news, gadgets and all other funny stuffs. Updated everyday.
Amazon Bedrock
GLM-4.7-Flash is a new member of the GLM 4.7 family and targets developers who want strong coding and reasoning performance in a model that is practical to...
Z.AI
The Hacker News thread on GLM-4.7-Flash (378 points, 135 comments, submitted by scrlk) captures community testing of the model in OpenCode running local 30B-A3B quantizations via llama.cpp on a 32 GB GPU with 128k context. One commenter reports that Qwen3-Coder had previously given the best results in their workflow of Early user issues documented in the thread include broken tool calling in OpenCode, rapid self-repetition (mitigated by raising llama.cpp's --dry-multiplier to 1.1 or higher), and spelling errors such as class or file-name characters being replaced with "1" or "AGENTS.md" being misread as "AGANTS.md". A later reply ann
Z.AI
AIbase reports that Zhipu AI officially open-sourced GLM-4.7-Flash in the early hours of January 20, 2026, framing it as a "Hybrid Thinking" model and the strongest competitor in the 30B specification class. The article confirms the 30B-A3B MoE design, where approximately 3 billion of 30 billion parameters activate per The excerpt lists the same vendor-published benchmark scores: 59.2 on SWE-bench Verified for code repair, 91.6 on AIME25 and 75.2 on GPQA for math and expert-level reasoning, 79.5 on τ²-Bench and 42.8 on BrowseComp for tool collaboration and agent scenarios. These figures are reported as surpassing Alibaba's Qwen3-30B-
This exact model name is also listed by 5 other providers.
