H Company Releases Holo4: Open-Weight Computer-Use Models That Click, Code and Call Tools Across Desktop, Web, Android and APIs
https://www.marktechpost.com/2026/09/29/h-company-releases-holo4-open-weight-computer-use-models-that-click-code-and-call-tools-across-desktop-web-android-and-apis/📌 【H Company】Holo4 開源:一套模型走遍桌面、網頁、Android 與 API
TL;DR：Holo4 是能點擊、寫程式、呼叫 MCP/API 工具的通用電腦操作模型,提供 Apache 2.0 開源權重可自架部署。
GUI agent 沒有畫面就當機,工具型 agent 遇到沒有 API 的軟體就卡住——H Company 想用同一套模型解決這兩種窘境。
🤔 為誰而做:跨平臺的通用電腦操作 agent
H Company 發表 Holo4,一系列給 AI agent 使用的通用電腦操作模型。它鎖定的痛點很明確:純 GUI 型 agent 沒有畫面就無法運作,純工具呼叫型 agent 遇到沒有 API 的應用程式就會卡住。Holo4 可以在桌面、網頁、Android、程式碼沙箱與商業 API 上運作,同一套模型、同一種呼叫方式,適用於每一種平臺。
🧩 架構:一套 vision-language 模型搭配開源 harness
Holo4 提供兩種尺寸:Holo4 27B(dense 架構)與 Holo4 35B-A3B(Mixture of Experts,啟用 3B 參數),兩者在 H Models API 上都支援 256K context。Holo4 27B 從 Qwen3.8-27B 微調而來,Holo4 35B-A3B 則建立在 Qwen3.6-35B-A3B 之上,兩者都搭配 H 公司開源的 hai-agents harness:harness 把螢幕截圖與工具回傳結果送給模型,再執行模型要求的點擊、輸入文字、程式碼與工具呼叫。
訓練資料方面,H 公司內部的 Agentic Task Factory 從文件、截圖與真實軟體中建立可驗證的任務環境,目前已產出約一萬個任務,其中網頁應用 4 千、MCP 伺服器 3 千、桌面與作業系統 3 千;任務要通過驗證器嚴格拒絕「差不多對」的結果,且 agent 必須透過真實介面完成才算解題成功。監督式微調(SFT)資料集共 127B token,約四分之三是成功的 agentic 軌跡,其中桌面佔 45%、網頁 14%、MCP 與 API 12%、行動裝置 3%。強化學習階段則以非同步線上 RL 訓練兩個 LoRA 專家,一個負責桌面與網頁,另一個負責終端機、MCP 與 API,兩者最後以相同權重合併,不再額外訓練。H 公司也依據 OSWorld 2.0 的失敗案例分析重建了 agent 迴圈,最大的改動是讓記憶在數百個步驟中保持可靠,以及在桌面機器上加入 shell。此外,H 公司也釋出以 NVIDIA Nemotron 3 Nano Omni 為基礎、透過 Nemotron Coalition 打造的 Holotron4 Nano,同樣的組合把 OSWorld 分數從基礎模型的 21.0% 拉高到 76.3%。
📊 短任務表現亮眼,長流程仍有落差
依 H 公司公布的評測表,Holo4 27B 在 OSWorld 上取得 85.2% 分數,每個任務成本 0.08 美元,其 Qwen3.8-27B 基礎模型則是 84.3% 分數、成本 0.22 美元;在 AndroidWorld 上,Holo4 27B 達到 85.1%。但在長流程任務上落差明顯:OSWorld 2.0 中,Holo4 27B 得分 61.7%、每任務成本 1.22 美元,而 Claude Opus 5.5(依 H 公司數據)得分 81.8%、成本 8.48 美元,H 公司也提醒不同廠商使用的 harness 與 effort 設定不同,跨廠商比較僅供參考方向。在 AutomationBench 上,Holo4 27B 整體得分 45.4%、每任務成本 0.05 美元,但需要留意的是,該評測 600 個公開任務中有 480 個落在 H 公司收集訓練資料的範圍內;在剩下 120 個完全held-out 的任務上,Holo4 27B 的分數是 49.3%。
授權方面,Holo4 35B-A3B 採 Apache 2.0 授權,可自行架設商用;Holo4 27B 則是 CC BY-NC 4.0,商用需透過 H Models API。定價上,Holo4 27B 每百萬 token 輸入 0.40 美元、輸出 3.00 美元,Holo4 35B-A3B 為 0.30 美元與 2.00 美元。API 相容 OpenAI 格式,端點為 https://api.hcompany.ai/v1;Hugging Face 上提供 BF16、FP8、NVFP4 與 4-bit GGUF 等多種格式的權重,並附上 vLLM 與 llama.cpp 的本機推論說明,H 公司表示未來幾天內還會釋出用於加速推論的 DSpark drafter checkpoint。所有的執行軌跡也公開在 trajectories.hcompany.ai 與 Hugging Face 上。
⚠️ 長流程與資料重疊是兩個要留意的地方
Holo4 在短任務、單步操作上的表現已相當接近甚至超越其基礎模型,但在需要維持數百步、跨情境記憶的長流程任務(OSWorld 2.0)上,與 Opus 5.5 之間仍有明顯差距,只是比較時使用的 harness 與 effort 設定不同,無法直接等同視之。另外,AutomationBench 的整體分數有一部分建立在與訓練資料重疊的任務上,完全 held-out 子集的分數更能反映真實泛化能力。
🎯 實務啟示
對想打造跨桌面、網頁、行動裝置與內部 API 的自動化 agent 團隊而言,Holo4 提供了一個可直接自架的開源選項(35B-A3B 為 Apache 2.0),搭配文件完整的 vLLM/llama.cpp 部署流程;其 Agentic Task Factory 的驗證式任務生成流程,也可以作為自建可驗證 agent 訓練資料的參考範本。
🔗 來源
- 標題:H Company Releases Holo4: Open-Weight Computer-Use Models That Click, Code and Call Tools Across Desktop, Web, Android and APIs
- 作者/機構:Michal Sutter(MarkTechPost)
- 連結:https://www.marktechpost.com/2026/09/29/h-company-releases-holo4-open-weight-computer-use-models-that-click-code-and-call-tools-across-desktop-web-android-and-apis/
#ComputerUse #OpenWeights #AIAgents #HCompany #Holo4 #MachineLearning #OpenSource #MCP #VisionLanguageModel #AgenticAI
原始資料 MarkTechPost · 收集於 2026-09-29
摘要原文
H Company has released Holo4 , a family of generalist computer-use models for AI agents. One set of weights clicks and types on screens. It also writes code and calls MCP or API tools. Holo4 ships in 2 sizes: Holo4 27B (dense) and Holo4 35B-A3B (Mixture of Experts, 3B active). Both serve a 256K context on the H Models API . Is it deployable? Yes. Holo4 35B-A3B ships Apache 2.0 weights for commercial self-hosting. Holo4 27B weights are CC BY-NC 4.0, so commercial use of 27B runs through the H Models API. Holo4 is a vision-language model for computer use. Holo4 27B is fine-tuned from Qwen3.8-27B . Holo4 35B-A3B is built on Qwen3.6-35B-A3B . Both pair with H’s open hai-agents harness . The harness sends screenshots and tool results to the model. It then executes the requested clicks, typing, code and tool calls. H company targets a known gap. GUI-only agents fail without a screen. Tool-calling agents stall when an application has no API. Holo4 runs on desktop, web, Android, code sandboxes and business APIs. It is the same model, called the same way, on every platform. Per H Company’s benchmark table , Holo4 27B scores 85.2% on OSWorld at $0.08 per task. Its Qwen3.8 27B base scores 84.3% at $0.22. On AndroidWorld, Holo4 27B reaches 85.1%. Long workflows show the remaining gap. On OSWorld 2.0, Holo4 27B scores 61.7% at $1.22 per task. Claude Opus 5.5 scores 81.8% at $8.48, per H’s figures. On AutomationBench, Holo4 27B scores 45.4% at $0.05 per task. It is important to note that frontier scores come from different harnesses and effort levels. Also, 480 of AutomationBench’s 600 public tasks sit in the split H Company collected training data from. On the 120 held-out tasks, Holo4 27B scores 49.3%. H Company publishes every trajectory at trajectories.hcompany.ai and on Hugging Face . Agentic Task Factory : H’s internal pipelines build environments and verifiable tasks from documentation, screenshots and real software. The factory has produced about 10,000 tasks: 4k web apps, 3k MCP servers, 3k desktop and OS. A task survives only if its verifier rejects near misses. An agent must also solve it through the real interface. Supervised fine-tuning : The SFT set holds 127B tokens. About three quarters are successful agentic trajectories: desktop 45%, web 14%, MCP and API 12%, mobile 3%. 2 RL experts, 1 merge : Asynchronous online RL trains 2 LoRA experts. One handles desktop and web. The other handles terminal, MCP and API. Both merge back with equal weight and no further training. Harness : H Company rebuilt its agent loop using OSWorld 2.0 failure analysis. The largest changes were reliable memory across hundreds of steps and a shell on the desktop machine. H Company also released Holotron4 Nano , built on NVIDIA’s Nemotron 3 Nano Omni through the Nemotron Coalition . The same combination lifts OSWorld from 21.0% to 76.3% over the base model. Holo4 27B costs $0.40 input and $3.00 output per 1M tokens. Holo4 35B-A3B costs $0.30 and $2.00. The API is OpenAI-compatible at https://api.hcompany.ai/v1 ; see the quickstart . Weights on the Hugging Face collection come in BF16, FP8, NVFP4 and 4-bit GGUF. H Company documents local inference with vLLM and llama.cpp. H says DSpark drafter checkpoints for faster inference arrive in the coming days. *OSWorld 2.0 figures from H Company’s newsroom table . Opus 5.5 ran at max effort in Anthropic’s harness. GPT-6 Astra ran at max effort on the 82-task offline subset. Harnesses differ, so treat cross-vendor rows as directional. Check out the Technical details , H Company’s HuggingFace Collection and Model API. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter . Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post H Company Releases Holo4: Open-Weight Computer-Use Models That Click, Code and Call Tools Across Desktop, Web, Android and APIs appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成