China OSS surge + GLM-5.2 on Huawei Ascend + MiniMax M3 + Qwen3.6/DeepSeek V4/Kimi K2.8 + Ornith-1.0 + Kimi Code CLI + GLM-5.3 vision poll + Tencent Hy3
Key Questions
What are the key specs and licensing details of GLM-5.2?
GLM-5.2 was trained on Huawei Ascend hardware at a reported $25M cost under an MIT license and is available quantized to 238GB. It supports 1-bit GGUF formats for consumer hardware use.
How does GLM-5.2 perform against DeepSeek V4 and Kimi K2.6?
In head-to-head benchmarks, GLM-5.2 achieves 62% on SWE Pro, matching or exceeding the other models in several coding tasks while offering strong open-weight accessibility.
What is Tencent Hy3 and how does it compare to GLM-5.2?
Hy3 is a 295B MoE model released under Apache 2.0 that is half the size of GLM-5.2, outperforming it on search and tool orchestration tasks but trailing in coding benchmarks.
What new coding tools were released alongside these models?
Moonshot AI released the open-source Kimi Code CLI agent supporting MCP, subagents, and video input, while Ornith-1.0 launched an agentic coding LLM family ranging from 9B to 397B parameters.
What do developer polls indicate for GLM-5.3?
Z.ai’s developer poll returned a unanimous request for vision capabilities in the upcoming GLM-5.3 release.
What share of tokens on OpenRouter comes from Chinese models?
OpenRouter data shows Chinese LLMs accounting for 61% of token usage, highlighting their growing adoption.
Are Chinese labs expected to continue open-sourcing frontier models?
HN discussions and reports note uncertainty, citing Qwen 3.7 remaining closed, Minimax M3 delays, and rumors that GLM and DeepSeek may restrict future releases.
How are enterprises using these Chinese open-weight models?
Coinbase’s CEO highlighted cost savings from Chinese models, while developers report daily use of GLM-5.2 via Hugging Face Inference Providers and fine-tuned Qwen3 variants achieving 84.7% accuracy in finance tasks at 13.8x lower cost.
China OSS surge continues with GLM-5.2 trained on Huawei Ascend at $25M cost, MIT license, quantized to 238GB. New head-to-head benchmark (GLM-5.2 vs DeepSeek V4 vs Kimi K2.6) shows 62% SWE Pro, with 1-bit GGUF for consumer hardware. Ornith-1.0 open-source agentic coding LLM family (9B-397B) released. Kimi Code CLI from Moonshot AI is a new open source coding agent supporting MCP, subagents, video input. OpenRouter shows 61% token share from Chinese LLMs. New: Z.ai developer poll shows unanimous demand for vision in GLM-5.3. Developer reports using GLM-5.2 daily in Claude Code via HF Inference Providers. Qwen3-235B fine-tuned for finance beats GPT/Claude with 84.7% accuracy and 13.8x cost reduction. New: Tencent's Hy3 under Apache 2.0 — 295B MoE, half size of GLM-5.2, wins on search/tool orchestration, removes geographic license restrictions. Uncertainty: HN discussion questions whether Chinese labs will keep open-sourcing — Qwen 3.7 not open, Minimax M3 delaying, rumors about GLM/Deepseek closing. New article 'The AI revolution will not be televised — it'll be quantized' reinforces commoditization narrative. Coinbase CEO also pitched Chinese open-weight models for cost savings.