--- title: "Qwen3.8-27B for Self-Hosted Computer Use" canonical: https://wavect.io/blog/qwen3-8-27b-self-hosted-computer-use-agents/ language: en description: "Qwen3.8-27B scores 84.3 on OSWorld-Verified. What a self-hosted computer-use agent can automate, what it must not touch, and how to evaluate it." image: "https://wavect.io/img/blog/headers/header_qwen3-8-27b-self-hosted-computer-use-agents.png" --- [**Back**](/blog/overview/) [![Kevin Riedl](/img/team/kevin.webp)](/team/kevin-riedl/) [Kevin Riedl](/team/kevin-riedl/) https://linkedin.com/in/wsdt 11 min read · 18 Aug 2026 Last reviewed August 18, 2026 [**Next**](/blog/hark-handoff-computer-use-agent-review/) # Qwen3.8-27B: self-hosted computer-use agents that keep screenshots in house TL;DR Qwen3.8-27B is a 27-billion-parameter dense multimodal model that Alibaba's Qwen team published on 14 August 2026 under Apache 2.0, with text, image and video input and 262,144 tokens of native context. Its published desktop-operation results lead the hosted models in the same table: 84.3 on OSWorld-Verified against 72.7 and 73.3 for the hosted comparisons, plus 81.9 on AndroidWorld. Browser work is weaker at 64.8 on WebArena-Verified, so roughly one task in three still fails there. Because screenshots cannot be redacted the way structured prompts can, this is the release that removes the data-transfer blocker from screen-automation projects, not the release that makes unattended autonomy safe. Serving it needs new measurements rather than dense-transformer estimates: 48 of the 64 layers use linear attention with a preallocated state cache sized by its own SGLang flags. Vendor benchmark tables are not comparable across harnesses, so pilot one process with 20 to 30 real tasks in fixture environments, plant prompt-injection attempts in editable fields, keep human confirmation in front of irreversible actions, and decide on cost per accepted task. **Qwen3.8-27B is the first open-weight model whose published desktop-operation score beats the hosted frontier models in its own comparison table.** That matters commercially for one reason: screen automation is the workload where sending data to someone else's API is hardest to defend. A screenshot carries whatever happened to be on the screen, including the customer record next to the invoice you wanted read. This article answers a narrow buying question: **can a self-hosted screen agent now do useful back-office work, and what has to be true before it touches a production system?** It is not a hardware sizing guide. For memory envelopes, quantization trade-offs and single-workstation pilots, use our [local agent model hardware and benchmark guide](/blog/muse-glimmer-30b-local-agent-guide/). If you would rather buy a hosted computer-use product than run one, read our [review of a research-preview computer-use agent](/blog/hark-handoff-computer-use-agent-review/). Keeping those three intents on separate pages is deliberate. ## What Qwen3.8-27B actually is Alibaba's Qwen team published the weights on 14 August 2026. The [official Qwen3.8-27B model card](https://huggingface.co/Qwen/Qwen3.8-27B) is the primary source for this section and for the benchmark table below. | Decision factor | Published position | What it means for a screen agent | | --- | --- | --- | | Size and type | 27B dense model, 64 layers, hidden size 5120 | Every parameter runs per token, so memory bandwidth sets your ceiling, not parameter count alone | | Layer layout | 16 repeats of 3 × (Gated DeltaNet → FFN) then 1 × (Gated Attention → FFN) | 48 linear-attention layers against 16 full-attention layers, which changes how long context consumes memory | | Inputs | Text, images and video in, text out | Screenshots and screen recordings are native inputs, not a bolted-on OCR step | | Context | 262,144 tokens native, extensible with YaRN scaling | Long GUI traces with many screenshots fit without aggressive summarisation | | Decoding | Multi-token prediction trained into the checkpoint | Speculative decoding is available without shipping a separate drafter model | | Licence | Apache 2.0 | Commercial use is intended, with notice and attribution duties on redistribution | ## The number that changes the build-versus-buy call OSWorld-Verified measures whether an agent finishes real tasks in a real desktop environment: open the file, change the setting, complete the form. It is the closest public proxy for the back-office work companies actually want automated. Qwen's published table puts a model you can run yourself ahead of the hosted models it lists. | Benchmark | Qwen3.8-27B | Qwen3.6-27B | Qwen3.7-Plus | Muse Glimmer-30B | Opus 4.6 Max | | --- | --- | --- | --- | --- | --- | | OSWorld-Verified (desktop operation) | **84.3** | 63.9 | 73.3 | 65.9 | 72.7 | | AndroidWorld (mobile operation) | **81.9** | 70.3 | 81.0 | not reported | 62.0 | | WebArena-Verified (browser tasks) | **64.8** | 48.8 | 55.3 | not reported | not reported | | Vision2Web | **62.9** | 45.0 | 42.1 | not reported | not reported | | SWE-bench Pro (repository work) | 61.7 | not reported | not reported | not reported | not reported | Read the WebArena row before the OSWorld row. A score of 64.8 on browser tasks means roughly one in three attempts fails on the surface where most commercial automation lives. The desktop result is genuinely strong. It is not a licence to point an agent at your ERP and leave the room. ## Why one vendor's table cannot settle a model choice Two vendors published different numbers for the same third-party model on the same benchmark within days of each other. Qwen reports 63.9 for its own previous generation, Qwen3.6-27B, on OSWorld-Verified. Meta's launch table for its 30B local model reports 75.6 for that same Qwen generation, as covered in our [Muse Glimmer guide](/blog/muse-glimmer-30b-local-agent-guide/). Neither number has to be wrong. Computer-use results depend on the scaffold: how many steps the agent may take, what screenshot resolution it sees, whether an accessibility tree is available, how tool errors are retried, which system prompt drives it. That is also why a vendor table cannot predict your result. The harness is part of the measurement, and in your deployment the harness is yours. **Our rule:** use published tables to decide what deserves a test, never to decide what ships. Freeze your own scaffold and measure again. The same discipline applies to money, which we model in [cost per token versus cost per accepted task](/blog/cost-per-token-vs-cost-per-task/). ## Screenshots are the payload you least want to export Text prompts can be filtered. A screen agent's input cannot be, in any complete sense. It sees the window it was told to operate plus everything around it: the neighbouring record, the open ticket, the internal dashboard, a colleague's name in a notification. Redaction before the call, the approach in [removing personal data before it reaches an LLM](/blog/pii-redaction-before-llm-prompts/), works on structured text. It does not survive contact with arbitrary pixels. That is why an open-weight model with credible desktop scores is a different proposition from a slightly better hosted one. Running inference inside your own network removes the transfer, the vendor retention question and one more sub-processor entry. It does not remove your obligations as the operator, and it adds the work of running the stack yourself. If the transfer is what blocks your risk assessment, this release removes the blocker. If the real blocker is that nobody on the team wants to operate GPUs, self-hosting does not help you. ## What the hybrid attention stack changes for your serving plan Capacity planning for a dense transformer does not carry over. Three quarters of the layers here are Gated DeltaNet layers, which hold a fixed-size recurrent state instead of a key-value cache that grows with every token. Long context therefore stops being a single linear memory curve, and estimates you memorised for a 27B transformer will mislead you in both directions. The [SGLang deployment cookbook for this model](https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8-27B) shows the consequence directly: the hybrid state cache gets its own sizing flags, including `--mamba-full-memory-ratio` and `--max-mamba-cache-size`, and the cookbook lists BF16, FP8 and NVFP4 W4A4 checkpoints with in-checkpoint multi-token prediction, targeting single-GPU serving on H200, RTX PRO 6000, RTX 5090 and DGX Spark class hardware. Four planning consequences follow, and each one is a measurement rather than a formula: - **Preallocation is a budget decision.** The state cache is reserved up front. Free memory at idle tells you very little about behaviour at your real concurrency. - **Screenshots are expensive tokens.** Image tokens dominate a GUI trace. Measure at the resolution and step count your agent will really use, not with a text prompt. - **Speculative decoding must be validated, not assumed.** Multi-token prediction helps single-stream latency most and can be close to neutral under heavy batching. - **Quantization is a quality experiment.** W4A4 changes what fits on one card. Whether it changes task success on your screens is a question only your own fixtures answer. If you are deciding how to serve this next to existing models, the platform patterns in [an inference stack built on vLLM and Triton](/blog/netflix-vllm-triton-inference-stack/) and the total-cost model in [self-hosting LLMs in the EU](/blog/self-hosting-llms-eu-cost/) apply unchanged. ## Where a self-hosted screen agent still loses The honest failure modes are not about capability. They are about what happens when the screen is adversarial or the action is irreversible. Screen agents read rendered pixels, so they cannot reliably tell a genuine interface element from user-generated content that merely looks like one. The [VPI-Bench study of visual prompt injection](https://arxiv.org/abs/2506.02456) reports that across 306 test cases on five widely used platforms, computer-use agents were deceived at rates of up to 51 percent and browser-use agents up to 100 percent on certain platforms, and that system-prompt defences produced only limited improvement. A comment field, a calendar invite title or a PDF footer is an instruction channel. | Scenario | Fit today | Why | | --- | --- | --- | | Read-only extraction from an internal system with no usable API | Strong pilot | Reversible, repetitive, and the data never leaves the network | | Legacy desktop software that will never get an API | Strong pilot | The interface is the only integration surface available | | Form filling with human confirmation before submit | Good pilot | The model drafts, a person accepts, the audit trail stays intact | | Cross-site browser workflows on public sites | Guarded only | A 64.8 browser score plus untrusted page content is a poor combination | | Payments, customer messaging, irreversible state changes | Not yet | Injection success rates above zero make unattended autonomy indefensible | | Stable high-volume process that already has an API | Wrong tool | A deterministic integration is cheaper, faster and easier to test | Before any agent gets credentials, put it through the controls in our [AI agent evaluation and sandbox checklist](/blog/ai-agent-eval-sandbox-security-checklist/). The structural rule is short: an agent that combines access to private data, exposure to untrusted content and a way to send data outward is exploitable. Break one of the three. ## A two-week evaluation that ends in a decision 1. **Pick one process and 20 to 30 real tasks.** Use last month's actual queue, including the messy cases and the ones a human escalated. 2. **Build fixture environments, not live systems.** Snapshot the application state so a run is repeatable and a failure is diagnosable. 3. **Freeze the scaffold.** Record checkpoint and quantization, runtime version, step limit, screenshot resolution, accessibility inputs, system prompt, tool set and retry policy. 4. **Measure accepted outcomes, not demos.** Track task completion, silent wrong actions, human correction minutes and steps per task. 5. **Measure the serving envelope.** Capture peak memory at target concurrency, prefill latency with image-heavy prompts, output speed and degradation under parallel sessions. 6. **Attack it on purpose.** Plant injected instructions in the fields your users can edit, then verify the agent refuses and logs the attempt. 7. **Compare against the boring baseline.** Run a deterministic script or an existing API on the same tasks. Sometimes the agent loses, and that is a useful result. 8. **Price the accepted task.** Include GPU capacity, engineering, monitoring, review time and failed runs, then compare with today's manual cost. A pass is not "the model ran". A pass is a completion rate above your process threshold, no silent destructive action in the adversarial set, a serving envelope that holds at real concurrency, and a cost per accepted task that beats the status quo. ## The compliance work self-hosting does not remove Keeping inference in house shortens the data-protection conversation. It does not end the AI Act conversation. According to [the European Commission's transparency guidance for Article 50](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act), those obligations apply from 2 August 2026, with a limited grace period until 2 December 2026 for the marking and detection duty covering systems already on the market before that date. Deployer duties are real but specific: informing people exposed to emotion recognition or biometric categorisation systems, and labelling deepfakes and AI-generated text on matters of public interest published without human editorial review. Two practical points for a screen agent. First, an internal back-office automation is often not an Article 50 system at all, so document the reasoning rather than assuming either answer. Second, substantially modifying an open-weight model can move you along the value chain toward provider duties, which is a question to settle with counsel before you fine-tune, not after. Our [Article 50 engineering checklist](/blog/eu-ai-act-article-50-checklist/) covers the implementation side, and [our trust center](/trust-center/) states plainly what we do and do not certify. This is operational guidance, not legal advice. ## Frequently asked questions ### What is Qwen3.8-27B? Qwen3.8-27B is a 27-billion-parameter dense multimodal model published by Alibaba's Qwen team on 14 August 2026 under Apache 2.0. It accepts text, images and video, produces text, supports 262,144 tokens of native context and uses a hybrid layer stack of Gated DeltaNet and Gated Attention layers. ### Can Qwen3.8-27B operate a computer or a browser? It is built for that. Qwen reports 84.3 on OSWorld-Verified for desktop operation, 81.9 on AndroidWorld and 64.8 on WebArena-Verified. Desktop results are strong, browser results still leave roughly one task in three failing, so scope the workload accordingly. ### Is a self-hosted computer-use agent better for GDPR than a hosted one? It removes the transfer and the vendor retention question, which is often the hardest part of a screen-automation assessment because screenshots cannot be reliably redacted. It does not remove your obligations as operator, and it adds the duty to run and secure the inference stack yourself. ### What hardware does Qwen3.8-27B need? The SGLang cookbook targets single-GPU serving on H200, RTX PRO 6000, RTX 5090 and DGX Spark class hardware, with BF16, FP8 and NVFP4 W4A4 checkpoints. Because 48 of the 64 layers use linear attention with a preallocated state cache, measure peak memory at your real context and concurrency instead of reusing dense-transformer estimates. ### What is the biggest risk in a screen agent pilot? Visual prompt injection. Agents read rendered pixels and cannot reliably separate interface from content, and VPI-Bench reports computer-use agents deceived at rates up to 51 percent with only limited benefit from system-prompt defences. Require human confirmation for irreversible actions and test with planted instructions. ### Can a company use Qwen3.8-27B commercially? The weights are published under Apache 2.0, which permits commercial use with licence, change-notice and attribution duties on redistribution and grants no trademark rights. Confirm the exact artifact and your distribution model with counsel, especially if you fine-tune. ## Final thoughts Screen automation has been the workload where the capability answer and the compliance answer pointed in opposite directions: the models good enough to operate a desktop were the ones you could not send your desktop to. Qwen3.8-27B closes that gap far enough to justify a real pilot on a real process. Treat the published scores as a reason to test, build fixtures instead of experimenting on live systems, attack your own agent before someone else does, and keep a person in front of every irreversible action. If it passes, the automation stays inside your network. If it does not, you found out in two weeks instead of after a rollout. ## You may also like.. [**Local agent models: hardware and benchmark guide** Size the memory envelope and read vendor benchmark tables with the right scepticism.](/blog/muse-glimmer-30b-local-agent-guide/) [**AI enablement vs generic AI consultancy** Compare a governed implementation with advice-only AI consulting.](/compare/ai-enablement-vs-generic-ai-consultancy/) Models and infrastructure ## Continue through this cluster Model selection, inference economics, local deployment, compression and serving architecture. [Start with the cornerstone**Self-Hosting LLMs in the EU: When Open Weights Actually Pay Off**](/blog/self-hosting-llms-eu-cost/) - [Netflix's vLLM and Triton Stack: 7 Production Lessons](/blog/netflix-vllm-triton-inference-stack/) - [Transformers.js Browser AI: When Local Inference Belongs in Your Product](/blog/transformers-js-browser-ai-guide/) - [How to Self-Host LiteLLM in Production: 2026 Guide](/blog/self-host-litellm-production-2026/) - [AI-Ready Company Wiki: Architecture and Build Guide](/blog/ai-ready-company-wiki/) - [Does Claude Watermark Text? The 2026 API Answer](/blog/claude-text-watermark-api-2026/) Inbox, without the noise ## Follow the work that matters to you Get a short email when we publish something new. Follow the whole blog or only the problems you care about. [**Back**](/blog/overview/) [![Kevin Riedl](/img/team/kevin.webp)](/team/kevin-riedl/) [Kevin Riedl](/team/kevin-riedl/) https://linkedin.com/in/wsdt 11 min read · 18 Aug 2026 Last reviewed August 18, 2026 [**Next**](/blog/hark-handoff-computer-use-agent-review/) New posts by email × × Get new posts by email A short email when we publish. Free, no tracking. ## Structured Data ```json { "@context": "https://schema.org", "@graph": [ { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ], "employee": [ { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } } ], "founder": [ { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } } ], "legalRepresentative": [ { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } } ], "name": "Wavect GmbH", "subjectOf": { "@id": "https://wavect.io/verified-claims.json#dataset", "@type": "Dataset", "creator": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] }, "description": "A machine-readable registry of quantitative and qualitative claims published by Wavect, with review dates, localized page appearances and public third-party citations where available.", "inLanguage": "en", "isAccessibleForFree": true, "license": "https://creativecommons.org/licenses/by/4.0/", "name": "Wavect verified publication claims", "url": "https://wavect.io/verified-claims.json" }, "url": "https://wavect.io/" }, { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "sameAs": [ "https://www.wikidata.org/wiki/Q139796365", "https://www.linkedin.com/in/wsdt", "https://github.com/wsdt" ], "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "sameAs": [ "https://www.wikidata.org/wiki/Q139796367", "https://www.linkedin.com/in/jocr77/", "https://github.com/jo-chris" ], "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/#website", "@type": "WebSite", "inLanguage": [ "en", "de", "es", "zh" ], "name": "Wavect", "potentialAction": { "@type": "SearchAction", "query-input": "required name=search_term_string", "target": { "@type": "EntryPoint", "urlTemplate": "https://wavect.io/search/?q={search_term_string}" } }, "publisher": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] }, "url": "https://wavect.io/" }, { "@id": "https://wavect.io/blog/qwen3-8-27b-self-hosted-computer-use-agents/#webpage", "@type": "WebPage", "dateModified": "2026-08-18", "inLanguage": "en", "isPartOf": { "@id": "https://wavect.io/#website", "@type": "WebSite" }, "lastReviewed": "2026-08-18", "url": "https://wavect.io/blog/qwen3-8-27b-self-hosted-computer-use-agents/" } ] } ``` ```json { "@context": "https://schema.org", "@type": "BlogPosting", "abstract": "Qwen3.8-27B is a 27-billion-parameter dense multimodal model that Alibaba's Qwen team published on 14 August 2026 under Apache 2.0, with text, image and video input and 262,144 tokens of native context. Its published desktop-operation results lead the hosted models in the same table: 84.3 on OSWorld-Verified against 72.7 and 73.3 for the hosted comparisons, plus 81.9 on AndroidWorld. Browser work is weaker at 64.8 on WebArena-Verified, so roughly one task in three still fails there. Because screenshots cannot be redacted the way structured prompts can, this is the release that removes the data-transfer blocker from screen-automation projects, not the release that makes unattended autonomy safe. Serving it needs new measurements rather than dense-transformer estimates: 48 of the 64 layers use linear attention with a preallocated state cache sized by its own SGLang flags. Vendor benchmark tables are not comparable across harnesses, so pilot one process with 20 to 30 real tasks in fixture environments, plant prompt-injection attempts in editable fields, keep human confirmation in front of irreversible actions, and decide on cost per accepted task.", "articleBody": " Blog overview/AI and agents/Models and infrastructure Qwen3.8-27B: self-hosted computer-use agents that keep screenshots in house TL;DR Qwen3.8-27B is a 27-billion-parameter dense multimodal model that Alibaba's Qwen team published on 14 August 2026 under Apache 2.0, with text, image and video input and 262,144 tokens of native context. Its published desktop-operation results lead the hosted models in the same table: 84.3 on OSWorld-Verified against 72.7 and 73.3 for the hosted comparisons, plus 81.9 on AndroidWorld. Browser work is weaker at 64.8 on WebArena-Verified, so roughly one task in three still fails there. Because screenshots cannot be redacted the way structured prompts can, this is the release that removes the data-transfer blocker from screen-automation projects, not the release that makes unattended autonomy safe. Serving it needs new measurements rather than dense-transformer estimates: 48 of the 64 layers use linear attention with a preallocated state cache sized by its own SGLang flags. Vendor benchmark tables are not comparable across harnesses, so pilot one process with 20 to 30 real tasks in fixture environments, plant prompt-injection attempts in editable fields, keep human confirmation in front of irreversible actions, and decide on cost per accepted task. Qwen3.8-27B is the first open-weight model whose published desktop-operation score beats the hosted frontier models in its own comparison table. That matters commercially for one reason: screen automation is the workload where sending data to someone else's API is hardest to defend. A screenshot carries whatever happened to be on the screen, including the customer record next to the invoice you wanted read. This article answers a narrow buying question: can a self-hosted screen agent now do useful back-office work, and what has to be true before it touches a production system? It is not a hardware sizing guide. For memory envelopes, quantization trade-offs and single-workstation pilots, use our local agent model hardware and benchmark guide. If you would rather buy a hosted computer-use product than run one, read our review of a research-preview computer-use agent. Keeping those three intents on separate pages is deliberate. What Qwen3.8-27B actually is Alibaba's Qwen team published the weights on 14 August 2026. The official Qwen3.8-27B model card is the primary source for this section and for the benchmark table below. Decision factorPublished positionWhat it means for a screen agent Size and type27B dense model, 64 layers, hidden size 5120Every parameter runs per token, so memory bandwidth sets your ceiling, not parameter count alone Layer layout16 repeats of 3 × (Gated DeltaNet → FFN) then 1 × (Gated Attention → FFN)48 linear-attention layers against 16 full-attention layers, which changes how long context consumes memory InputsText, images and video in, text outScreenshots and screen recordings are native inputs, not a bolted-on OCR step Context262,144 tokens native, extensible with YaRN scalingLong GUI traces with many screenshots fit without aggressive summarisation DecodingMulti-token prediction trained into the checkpointSpeculative decoding is available without shipping a separate drafter model LicenceApache 2.0Commercial use is intended, with notice and attribution duties on redistribution The number that changes the build-versus-buy call OSWorld-Verified measures whether an agent finishes real tasks in a real desktop environment: open the file, change the setting, complete the form. It is the closest public proxy for the back-office work companies actually want automated. Qwen's published table puts a model you can run yourself ahead of the hosted models it lists. BenchmarkQwen3.8-27BQwen3.6-27BQwen3.7-PlusMuse Glimmer-30BOpus 4.6 Max OSWorld-Verified (desktop operation)84.363.973.365.972.7 AndroidWorld (mobile operation)81.970.381.0not reported62.0 WebArena-Verified (browser tasks)64.848.855.3not reportednot reported Vision2Web62.945.042.1not reportednot reported SWE-bench Pro (repository work)61.7not reportednot reportednot reportednot reported Read the WebArena row before the OSWorld row. A score of 64.8 on browser tasks means roughly one in three attempts fails on the surface where most commercial automation lives. The desktop result is genuinely strong. It is not a licence to point an agent at your ERP and leave the room. Why one vendor's table cannot settle a model choice Two vendors published different numbers for the same third-party model on the same benchmark within days of each other. Qwen reports 63.9 for its own previous generation, Qwen3.6-27B, on OSWorld-Verified. Meta's launch table for its 30B local model reports 75.6 for that same Qwen generation, as covered in our Muse Glimmer guide. Neither number has to be wrong. Computer-use results depend on the scaffold: how many steps the agent may take, what screenshot resolution it sees, whether an accessibility tree is available, how tool errors are retried,", "articleSection": "Engineering", "author": { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "name": "Kevin Riedl", "sameAs": [ "https://www.wikidata.org/wiki/Q139796365", "https://www.linkedin.com/in/wsdt", "https://github.com/wsdt" ], "url": "https://wavect.io/team/kevin-riedl/" }, "citation": [ { "@type": "WebPage", "name": "official Qwen3.8-27B model card", "url": "https://huggingface.co/Qwen/Qwen3.8-27B" }, { "@type": "WebPage", "name": "SGLang deployment cookbook for this model", "url": "https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8-27B" }, { "@type": "WebPage", "name": "VPI-Bench study of visual prompt injection", "url": "https://arxiv.org/abs/2506.02456" }, { "@type": "WebPage", "name": "the European Commission's transparency guidance for Article 50", "url": "https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act" } ], "dateModified": "2026-08-18", "datePublished": "2026-08-18", "description": "Qwen3.8-27B is a 27-billion-parameter dense multimodal model that Alibaba's Qwen team published on 14 August 2026 under Apache 2.0, with text, image and video input and 262,144 tokens of native context. Its published desktop-operation results lead the hosted models in the same table: 84.3 on OSWorld-Verified against 72.7 and 73.3 for the hosted comparisons, plus 81.9 on AndroidWorld. Browser work is weaker at 64.8 on WebArena-Verified, so roughly one task in three still fails there. Because screenshots cannot be redacted the way structured prompts can, this is the release that removes the data-transfer blocker from screen-automation projects, not the release that makes unattended autonomy safe. Serving it needs new measurements rather than dense-transformer estimates: 48 of the 64 layers use linear attention with a preallocated state cache sized by its own SGLang flags. Vendor benchmark tables are not comparable across harnesses, so pilot one process with 20 to 30 real tasks in fixture environments, plant prompt-injection attempts in editable fields, keep human confirmation in front of irreversible actions, and decide on cost per accepted task.", "headline": "Qwen3.8-27B: Self-Hosted Computer-Use Agents Without Exporting Screenshots", "image": "https://wavect.io/img/blog/headers/header_qwen3-8-27b-self-hosted-computer-use-agents.svg", "inLanguage": "en", "keywords": "Computer Use, Local AI", "mainEntityOfPage": { "@id": "https://wavect.io/blog/qwen3-8-27b-self-hosted-computer-use-agents/", "@type": "WebPage" }, "publisher": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] }, "url": "https://wavect.io/blog/qwen3-8-27b-self-hosted-computer-use-agents/", "wordCount": 2629 } ``` ```json { "@context": "https://schema.org", "@type": "BreadcrumbList", "itemListElement": [ { "@type": "ListItem", "item": "https://wavect.io/", "name": "Home", "position": 1 }, { "@type": "ListItem", "item": "https://wavect.io/blog/overview/", "name": "Blog overview", "position": 2 }, { "@type": "ListItem", "item": "https://wavect.io/blog/topics/ai-agents/", "name": "AI and agents", "position": 3 }, { "@type": "ListItem", "item": "https://wavect.io/blog/clusters/models-infrastructure/", "name": "Models and infrastructure", "position": 4 }, { "@type": "ListItem", "item": "https://wavect.io/blog/qwen3-8-27b-self-hosted-computer-use-agents/", "name": "Qwen3.8-27B for Self-Hosted Computer Use | ", "position": 5 } ] } ``` ```json { "@context": "https://schema.org", "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "Qwen3.8-27B is a 27-billion-parameter dense multimodal model published by Alibaba's Qwen team on 14 August 2026 under Apache 2.0. It accepts text, images and video, produces text, supports 262,144 tokens of native context and uses a hybrid layer stack of Gated DeltaNet and Gated Attention layers." }, "name": "What is Qwen3.8-27B?" }, { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "It is built for that. Qwen reports 84.3 on OSWorld-Verified for desktop operation, 81.9 on AndroidWorld and 64.8 on WebArena-Verified. Desktop results are strong, browser results still leave roughly one task in three failing, so scope the workload accordingly." }, "name": "Can Qwen3.8-27B operate a computer or a browser?" }, { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "It removes the transfer and the vendor retention question, which is often the hardest part of a screen-automation assessment because screenshots cannot be reliably redacted. It does not remove your obligations as operator, and it adds the duty to run and secure the inference stack yourself." }, "name": "Is a self-hosted computer-use agent better for GDPR than a hosted one?" }, { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "The SGLang cookbook targets single-GPU serving on H200, RTX PRO 6000, RTX 5090 and DGX Spark class hardware, with BF16, FP8 and NVFP4 W4A4 checkpoints. Because 48 of the 64 layers use linear attention with a preallocated state cache, measure peak memory at your real context and concurrency instead of reusing dense-transformer estimates." }, "name": "What hardware does Qwen3.8-27B need?" }, { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "Visual prompt injection. Agents read rendered pixels and cannot reliably separate interface from content, and VPI-Bench reports computer-use agents deceived at rates up to 51 percent with only limited benefit from system-prompt defences. Require human confirmation for irreversible actions and test with planted instructions." }, "name": "What is the biggest risk in a screen agent pilot?" }, { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "The weights are published under Apache 2.0, which permits commercial use with licence, change-notice and attribution duties on redistribution and grants no trademark rights. Confirm the exact artifact and your distribution model with counsel, especially if you fine-tune." }, "name": "Can a company use Qwen3.8-27B commercially?" } ] } ```