One-line: From-scratch 7.39B dense math + agentic-search model from Zhongguancun Academy/ZIAI — GAIA text-only 42.52 with up to 64 web search-and-read steps, 256K context, MIT license, fully open training recipe.
⚠ Surfaced via web-search/HF trending-papers (below the API streams’ download cutoff at 406 downloads, createdAt 09-07 just outside the 36h window). All numbers are lab-reported from the tech report (arXiv 2609.13356) — no independent third-party eval yet; custom ZgcmForCausalLM architecture requires trust_remote_code=True, so llama.cpp/Ollama support is not a given.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
not published
HumanEval+
—
not published
MBPP
—
not published
LiveCodeBench
—
not published
Aider polyglot
—
not published
SWE-bench Verified
—
not published
SWE-bench Multilingual
—
not published
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
not published
τ-Bench
—
not published
ToolACE
—
not published
GAIA
42.52 (text-only)
lab-reported (tech report)
WebWalkerQA
63.09
lab-reported (tech report)
BrowseComp
19.43
lab-reported (tech report)
Binary Function Search
62.00 (31/50 Ghidra tool)
lab-reported (tech report)
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
not published
GPQA Diamond
—
not published
MATH-500
97.13
lab-reported
AIME 2024
80.62
lab-reported
AIME 2025
73.33
lab-reported
AIME 2026
75.00
lab-reported
HMMT 2025
70.42
lab-reported
Efficiency
Active params: 7.39B dense (32 layers, hidden 4096; 27 gated sliding-window GQA layers + 5 global layers)
Context length: 262,144 tokens (256K)
VRAM (fp16 rough): ~15 GB — needs quant (Q4 ≈ 4–5 GB) for consumer GPUs and Pi-class devices; ⚠ custom architecture means GGUF availability unverified
What makes it notable
Strongest published agentic-search numbers at 7B scale: GAIA text-only 42.52 and WebWalkerQA 63.09 with up to 64 web-search steps is well above typical 7–8B VL/agent models, and the MATH-500 97.13 / AIME 2026 75.00 pair tops the report’s 7–8B comparison table (DeepSeek-R1-0528-Qwen3-8B, MiniCPM4.1-8B, Qwen3-8B, Olmo 3 7B Think).
Best average rank across 14 reasoning benchmarks among seven 7–8B models — a genuine from-scratch contender, not a Qwen/Llama fine-tune.
Fully open recipe: MIT weights, dataset, and complete training code (pretrain → 16K/64K/256K mid-training → SFT/RL with MDP-formulated interaction traces).
For agent loops: tool-use strength is web search/read + Ghidra-style binary function search; no SWE-bench or BFCL numbers published, so repo-coding suitability is unverified — treat as a research/math-agent model until independent evals appear.