--- title: "Qwen Models: Benchmarks, Licences and Self-Hosting" description: "A senior lab's read on the Qwen family: Qwen3.8-27B under Apache 2.0, Flash-Next at 180B sparse, and when self-hosting actually pays." source: "https://www.kensink.com/models/qwen/" canonical: "https://www.kensink.com/models/qwen/" --- ★ Qwen LLM Models 8-week engagement QWEN · ALIBABA · OPEN WEIGHTS # The open-weight family everything else forks. Qwen ships frontier-adjacent models under Apache 2.0 at sizes from 0.6B to 180B, with vision in the base checkpoint rather than bolted on. It is the most-downloaded and most-derived family in open weights, and the answer we reach for when data cannot leave the network. Open weights Self-hosted Eval pipelines [Start a conversation →](https://www.kensink.com/contact) [All llm models →](https://www.kensink.com/models) Cycle 8 weeks · fixed price Stack Qwen, hosted or self-hosted Output Production code + eval suite Handoff Full source ownership \[THE SHORT VERSION\] ## The licence is doing as much work as the benchmarks. Qwen3.8-27B beats Claude Opus 4.6 Max on agentic coding and computer use in Alibaba's own table, which is remarkable at 27B. The part that decides projects is Apache 2.0: commercial use with no negotiation, no field-of-use restriction, and no vendor able to revise the terms later. We integrate the hosted API and the self-hosted weights behind one abstraction and route on privacy, latency and volume rather than on ideology. When it fits - Data that cannot leave your network or your jurisdiction - Sustained high volume where per-token billing stops making sense - Latency floors that a network round trip cannot meet - Anywhere a commercial licence review would outlast the build When it does not - Low or spiky volume, where the hosted tier at $0.14 per million is cheaper than your GPUs - Teams without the appetite to own filtering, monitoring and audit themselves \[HOW WE BUILD IT\] ## How we build with Qwen. 01 ### Scope and fit We decide where Qwen earns its place in your system, and where a simpler tool wins. No resume-driven architecture. 02 ### Build on a tested foundation We integrate Qwen against a foundation we trust: typed code, CI, and observability from the first commit. Boring infrastructure, modern surface. 03 ### Eval before launch An eval suite proves the build behaves before it reaches a user. We measure, then ship. 04 ### Handoff with ownership Your team gets the code, the tests, and a runbook. No lock-in to us or to a vendor framework. \[ WHAT YOU GET \] ## What the engagement leaves behind. Senior Engineers who have shipped this before 100% Source ownership at handoff Eval-first Tested before it ships 0 Framework lock-in \[MODELS + VENDORS\] ## The open-weight family everything else forks. Qwen ships frontier-adjacent models under Apache 2.0 and a community licence, at sizes from 0.6B to 180B, with vision in the base model rather than bolted on. Count the quantisations, abliterations and fine-tunes on Hugging Face and it is the most-derived family in open weights by a wide margin. We integrate the hosted API and self-hosted weights behind the same abstraction. [ Flagship Alibaba 5 Aug 2026 ### Qwen3.8-27B Qwen/Qwen3.8-27B Dense multimodal, self-hostable Licence Apache 2.0 Params 27.8B dense - The most-liked model on Hugging Face after FLUX, at 14.2k likes and 6.4M downloads - 262K context native, extensible to 1M. Text, images and video in - Apache 2.0, so commercial use needs no negotiation Read the technical brief ](https://www.kensink.com/models/qwen/qwen3-8-27b)[ Flagship Alibaba 24 Aug 2026 ### Qwen3.8-Flash-Next Qwen/Qwen3.8-Flash-Next Sparse multimodal, 6B active Licence qwen-community-1.0 Params 180B / 6B active - 180B on disk, 6B activated per token. Frontier scores at small-model inference cost - Beats the 27B on every published benchmark, and Claude Opus on several - Community licence rather than Apache 2.0, so read the terms before you ship Read the technical brief ](https://www.kensink.com/models/qwen/qwen3-8-flash-next) Current Alibaba 3 Aug 2026 ### Qwen3.8-Max qwen3.8-max Hosted flagship, closed weights Input $2 / 1M Output $6 / 1M - The hosted tier with no public weights. Cached input at $0.25 per million - A fifth of Claude Opus 5 on input, half of GPT-5.6 Sol - Singapore endpoint pricing. The Beijing endpoint runs 60 to 70% cheaper In the routing table · brief not yet published Current Alibaba 26 Aug 2026 ### Qwen3.8-Flash qwen3.8-flash Hosted cheap tier Input $0.14 / 1M Output $0.42 / 1M - Flat rate across the full 1M context, with no long-context surcharge - Roughly a seventieth of GPT-6 Astra on input - The tier we route classification, routing and extraction to In the routing table · brief not yet published Current Alibaba 2026 ### Qwen3-VL-8B-Instruct Qwen/Qwen3-VL-8B-Instruct Vision-language, small Downloads 14.5M / month - One of the most-downloaded vision-language models on Hugging Face - 8B, so it fits on a single accelerator for document and screen work - The open-weight default when images cannot leave your network In the routing table · brief not yet published Previous Alibaba 2025 ### Qwen3-8B Qwen/Qwen3-8B General purpose, small Downloads 13.0M / month - The workhorse size. Still 13M downloads a month a year after release - Fits comfortably on one GPU and quantises well - Where most self-hosted pilots start before they size up In the routing table · brief not yet published Current Alibaba 2025 ### Qwen3-0.6B Qwen/Qwen3-0.6B Edge and on-device Downloads 21.3M / month - 21M downloads a month, more than any frontier open model - Small enough for on-device, browser and edge-runtime inference - Right for routing and classification where latency beats judgement In the routing table · brief not yet published Current Alibaba 2026 ### Qwen3-ASR-1.7B Qwen/Qwen3-ASR-1.7B Speech to text Downloads 3.3M / month - Open-weight transcription at a size you can actually self-host - The alternative to Whisper when audio cannot leave your infrastructure - Covered alongside the hosted options on our voice page In the routing table · brief not yet published Current Alibaba 2025 ### Qwen3-Reranker-4B Qwen/Qwen3-Reranker-4B Reranking for RAG Downloads 2.4M / month - Reranking is the cheapest quality win available in a RAG pipeline - Sits after retrieval and before generation, and fixes more than a better embedder does - A 0.6B variant exists for latency-bound paths In the routing table · brief not yet published \[METHODOLOGY · K-FRAMEWORK\] ## Integrated through the K-Framework. Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship. [Read the K-Framework](https://www.kensink.com/k-framework) 01 ### Foundations Direct API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres. 02 ### Amplification An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo. 03 ### Judgment Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots. \[OBSERVABILITY\] ## Observability your team can read. A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket. Instrumented ### Cost per call Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters. Instrumented ### Latency p50 / p95 / p99 Real distributions, not averages. We know which routes are slow, and why. Instrumented ### Eval pass rates The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast. Instrumented ### Prompt + completion logs PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window. Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics. \[COMMON QUESTIONS\] ## Questions we get asked. Is Qwen actually free to use commercially? Qwen3.8-27B is Apache 2.0, which grants commercial use and a patent licence with no field-of-use restriction and no user threshold. That is not true of every tier: Qwen3.8-Flash-Next ships under qwen-community-1.0. Check the licence per checkpoint rather than per family, because they differ inside the same release window. How does Qwen compare to Claude and GPT? In Alibaba's own table, Qwen3.8-27B beats Claude Opus 4.6 Max on SWE-bench Pro (61.7 against 53.4) and OSWorld computer use (84.3 against 72.7), and trails on terminal coding. Those are vendor-selected comparisons against a model that is no longer the current Claude tier. Read them as evidence that a 27B open model is in the conversation, not as a verdict. What hardware do we need to self-host it? About 56GB of weights in BF16 for the 27B, so a single 80GB accelerator with headroom for context, or considerably less quantised. Because it is dense there is no expert routing to plan around. Size against your p95 context length and concurrency rather than a single-request demo. Should we self-host or use the hosted API? Self-host when data cannot leave your boundary, when you need the latency floor of local inference, or when sustained volume beats the GPU bill. Otherwise the hosted tier usually wins: Qwen3.8-Flash is $0.14 input and $0.42 output per million, which is hard to beat once you count engineering time and on-call. Do the published benchmarks apply to the quantised builds? No, and this is the trap we see most often. The numbers are BF16 and almost nobody deploys BF16. A 4-bit GGUF is a different model whose losses concentrate in long-context and tool-calling behaviour rather than in short prompts. Evaluate the exact artefact going to production. Share[](https://twitter.com/intent/tweet?url=https%3A%2F%2Fwww.kensink.com%2Fmodels%2Fqwen%2F&text=Qwen%20%C2%B7%20LLM%20Models)[](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fwww.kensink.com%2Fmodels%2Fqwen%2F) [View .md](https://www.kensink.com/models/qwen.md) \[RELATED\] ## Worth a look next. [ LLM Models GLM Read more ](https://www.kensink.com/models/glm)[ LLM Models DeepSeek Read more ](https://www.kensink.com/models/deepseek)[ LLM Models Llama Read more ](https://www.kensink.com/models/llama) APPLIED K-FRAMEWORK ## Bring the problem. We’ll bring the build. Senior engineers, eval suite at handoff, full source ownership. Sprint, program, or ongoing. We shape the engagement to the work. [Start a conversation →](https://www.kensink.com/contact) [Read the K-Framework](https://www.kensink.com/k-framework)