Open 7.39B dense model for math reasoning and agentic search, with a 256K context and the full training pipeline released.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
ZGCM-1 is a 7.39B-parameter dense decoder-only language model trained from scratch by Zhongguancun Academy and Zhongguancun Institute of Artificial Intelligence. It targets mathematical reasoning and tool-assisted search, supports a 256K-token context, and offers both thinking and direct-response modes in one model.
The architecture uses 32 layers with 27 gated sliding-window attention layers (128-token window) and 5 global layers, with grouped-query attention. Training covered about 4.19T pretraining tokens with FP8 precision and the Muon optimizer, about 600B mid-training tokens extending context from 16K to 64K to 256K, and supervised fine-tuning on general and agentic data. The report also explores mixed RL for math and code.
Reported results for the 256K SFT checkpoint in thinking mode include 97.13% on MATH-500, 75.00% on AIME 2026 and 70.42% on HMMT 2025. Agentic search results include 63.09% on WebWalkerQA, 19.43% on BrowseComp and 42.52% on text-only GAIA. These are the authors' own numbers. The comparison set is other 7B–8B models, some of them older generations, so the ranking claims should be read with that in mind.
The GitHub repository provides data-processing, pretraining, mid-training, SFT and RL workflows under an MIT license. Weights and data are published on Hugging Face.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Small 7B open model for mathematical reasoning and agentic search, with the full training pipeline released. The creator notes it is compared against older model generations, which is somewhat cherry-picked, though it ranks well on competitive math benchmarks.
Closest neighbours by meaning, not just by category.
Model · Meituan
Meituan's open-weights 1.6T-parameter mixture-of-experts LLM (48B active), aimed at agentic coding and reasoning tasks.
Model · DeepSeek
DeepSeek's open-weights reasoning LLM (671B MoE, 37B active) trained with large-scale reinforcement learning.
Model
35B Mixture-of-Experts open-weights model from InternScience for long-horizon agentic search, engineering and tool use.
Model
Open-weights 35B-A3B agentic model post-trained from Qwen3.6-35B-A3B for search, coding and scientific research.
Model · OpenAI
OpenAI's open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, released under Apache 2.0.
Model · OpenBMB
1B-parameter dense open-weights language model from OpenBMB for on-device use, with hybrid thinking and 128K context.