--- title: "360 open-sources a 7B model that swallows a whole book in one context" date: 2026-10-08 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-360zhinao-long-context language: en --- # 360 open-sources a 7B model that swallows a whole book in one context > Qihoo 360's 360Zhinao3 open model pairs a 7B core with a 360,000-token window and a long chain-of-thought variant. Most open models forget the start of a long document by the time they reach the end. A security company best known for antivirus just open-sourced a 7B model that can hold an entire book in its head at once. The bet is that **context length (长上下文)**, not just raw parameter count, is where Chinese open models can quietly lead. ## The 360,000-token window The model family is **360Zhinao (360智脑)**, from **Qihoo 360 (三六零)**, the Beijing/Shanghai security group formally named **360 Security Technology Inc. (三六零安全科技股份有限公司)**. Its headline number is a **360,000-token context (360K tokens)**. In plain terms, that is roughly half a million Chinese characters — close to a short monograph, a full legal contract stack, or a year of chat logs in a single prompt. - The long window first appeared with 360Zhinao-7B v1.0 in April 2024, offered at 4K, 32K, and 360K lengths. - At launch, the team claimed 360K was the longest context among domestic open models of its time. - The newest generation keeps that ceiling while improving the base. ## What you actually get In April 2025 the lab released **360Zhinao3 (360智脑3)**, a refresh of the 7B line under the **Apache-2.0** license, free for commercial use. - 360Zhinao3-7B — the base, continued pretraining on **700B high-quality tokens** atop the previous 360Zhinao2-7B. - 360Zhinao3-7B-Instruct — the chat-tuned version. - 360Zhinao3-7B-O1.5 — a **long chain-of-thought (长思维链)** model for multi-step reasoning. The O1.5 variant targets exactly the problems small models usually flunk: competition math and multi-hop science. Third-party model libraries report it at **AIME24 54.2** and **GPQA Diamond 40**, though those are repository-listed figures rather than our own runs. Two smaller siblings also topped Chinese retrieval leaderboards: 360Zhinao-search and 360Zhinao-1.8B-Reranking ranked first on **C-MTEB** Retrieval and Reranking in May 2024. For a company whose business is sifting threats from noise, search and ranking are the natural home turf. ## Built like a lab rat, not a flagship 360Zhinao is deliberately mid-sized. The training recipe reads like efficient engineering rather than a show of force. - 360Zhinao2-7B was trained on about **10.1 trillion tokens (10.1T tokens)** across two stages, then 360Zhinao3 added 700B more. - The 7B size means it runs on a single workstation-class GPU, not a data-center pod. - The O1.5 long-CoT variant is the "think longer" answer to reasoning models, but at a fraction of the cost of frontier systems. Qihoo 360's AI lineage goes back further: its 360GPT-S2-V8 was, by the company's account, the first domestic model to pass the **China Academy of Information and Communications Technology (中国信通院, CAICT)** capability appraisal, back in June 2023. ## Why a security firm is shipping models This is the part that surprises outsiders. **Zhou Hongyi (周鸿祎)**, the company's chairman, has pushed 360 to treat large models (大模型) as the next platform, not just a feature. The logic is defensive. A security company lives on: - ingesting massive, mixed-format logs and documents; - retrieving the one relevant signal from noise; - reasoning about intent over long sequences. A long-context, open, cheaply runnable model fits that workload better than a closed API with a short window and per-token billing. Open-sourcing also seeds an ecosystem that, in 360's calculus, strengthens the domestic stack it depends on. The parent is publicly listed, ticker **601360** on the Shanghai exchange. Its late-September 2026 market value was about **RMB 61.4 billion (roughly US$8.6 billion / HK$66.9 billion)** — a security incumbent, not a model pure-play, bankrolling open research. ## Where the long window earns its keep A 360K context is not a party trick. It is what lets the model treat a whole document set as a single object rather than a chopped-up prompt. - A security team can drop a quarter of threat logs and ask for the thread running through all of them, not just the last page. - A legal or compliance user can question an entire contract pack without splitting it into chunks that quietly lose cross-references. - A developer can paste a large code slice and get answers that respect relationships across files. That is the wedge a security-first company cares about: reasoning over long, messy, real-world sequences instead of tidy demo prompts. It is also why 360 pairs the long window with retrieval models — 360Zhinao-search and the 1.8B-Reranking model that each hit first place on the **C-MTEB** leaderboard in May 2024 handle the "find the needle" step, while the 360K context handles the "now read the whole haystack" step. ## What readers can do now - **Test the 360K window directly**: paste a long report and ask for cross-references near the start and end; most 8K models will silently drop the opening. - **Use the O1.5 variant for reasoning, not the base**: the long chain-of-thought model is the one built for math and science chains. - **Check the Apache-2.0 terms before shipping**: free commercial use is allowed, but validate on your own data and keep the license headers. ## Honest limitations - The 360K context is officially supported, but very long prompts cost memory and latency; real throughput varies by hardware. - AIME24 and GPQA scores for O1.5 come from a model-library listing, not an independent benchmark we ran. - "Longest domestic open context at launch" was a 2024 claim; the field moves fast and others have since pushed further. - A 7B model, even with a long window, will trail larger systems on breadth and factual density — context is not a substitute for capacity. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-360zhinao-long-context Free to quote with attribution and a link to the original.