--- title: Qwen 3.8 Flash Next vs Qwen 3.8 27B, which one to run and when | daily.dev description: Choose 27B for local/private runs and Flash Next for hosted coding agents or extreme long-context tasks; plan for VRAM, quantization, and runtime. canonical: https://daily.dev/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/ og:type: article og:url: https://daily.dev/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/ og:title: Qwen 3.8 Flash Next vs Qwen 3.8 27B, which one to run and when | daily.dev og:description: Choose 27B for local/private runs and Flash Next for hosted coding agents or extreme long-context tasks; plan for VRAM, quantization, and runtime. og:image: https://media.daily.dev/image/upload/s--oPUMc6tK--/f_auto,q_auto/v1/recruiter-landing/6ae96e8df0eee09effb2869c_1793684818134_65e92b395f?_a=BAMAMiB80 og:site_name: daily.dev og:locale: en_US article:published_time: 2026-11-03 article:modified_time: 2026-11-03T06:16:16.589Z article:author: Kevin Nguyen twitter:card: summary_large_image twitter:site: @dailydotdev twitter:creator: @dailydotdev twitter:title: Qwen 3.8 Flash Next vs Qwen 3.8 27B, which one to run and when | daily.dev twitter:description: Choose 27B for local/private runs and Flash Next for hosted coding agents or extreme long-context tasks; plan for VRAM, quantization, and runtime. twitter:image: https://media.daily.dev/image/upload/s--oPUMc6tK--/f_auto,q_auto/v1/recruiter-landing/6ae96e8df0eee09effb2869c_1793684818134_65e92b395f?_a=BAMAMiB80 --- **I’d choose [Qwen 3.8 27B](https://huggingface.co/Qwen/Qwen3.8-27B) for local use and Flash Next for hosted coding agents or very long prompts.** On a 24 GB GPU, start with 27B at 4-bit - about 16–17 GB of weights - and leave room for context and runtime memory. The catch? _Fewer active parameters doesn’t mean less memory._ Flash Next uses 6B parameters per token, but its full BF16 weights take about 335–360 GB. **Quick Comparison** | What I’d compare | [Qwen 3.8 Flash Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) | [Qwen](https://qwen.ai/research) 3.8 27B | | --- | --- | --- | | Architecture | Sparse mixture of experts | Dense | | Best fit | Hosted agents, multi-step coding, long prompts | Local chat, coding, and private document work | | Memory | Server-class capacity or heavy offloading | About 16–17 GB of 4-bit weights | | Coding | Leads on agent tasks | Close on single-turn code generation | | Context and documents | Better suited to extreme context lengths | Same stated context limit; similar document scores | | Input support | Text, images, video | Text, images, video | | Hosted price per 1M tokens | $0.15 input / $0.50 output | $0.265 input / $0.65 output | | License | [Qwen Community License 1.0](https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE) | [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0) | | Tools | [Ollama](https://ollama.com/), [llama.cpp](https://github.com/ggml-org/llama.cpp), [vLLM](https://vllm.ai/); check model support | Ollama, [LM Studio](https://lmstudio.ai/), llama.cpp, vLLM | My rule: check memory and runtime support before chasing benchmark scores. For either model, I’d test the context length, tool calls, and document retrieval your workload needs - not just whether the weights load. ::: @figure {Qwen 3.8 Flash Next vs 27B: Which Should You Run?} ## Local speed and memory needs **Download size is not your runtime memory budget.** You also need room for loaded weights, the KV cache, and runtime buffers. ### Consumer GPUs and the 24 GB limit For Qwen3.8-27B, official FP8 weights take up about 30.9 GB. The 4-bit Q4\_K\_M weights take up about 15.8–17 GB[\[1\]](https://featherless.ai/blog/glm-5-3-flash-vs-qwen-3-8-flash-next-vs-qwen-3-8-27b)[\[8\]](https://blog.laozhang.ai/ko/posts/qwen3-8-flash-next-vs-qwen3-8-27b). Fitting Qwen3.8-27B on a 24 GB GPU requires 4-bit quantization, but **long context can still exceed that limit**[\[1\]](https://featherless.ai/blog/glm-5-3-flash-vs-qwen-3-8-flash-next-vs-qwen-3-8-27b)[\[8\]](https://blog.laozhang.ai/ko/posts/qwen3-8-flash-next-vs-qwen3-8-27b). At 262,144 tokens, the 27B model’s KV cache alone uses about 16.0 GB[\[6\]](https://kgptalkie.com/tutorials/llm-benchmarking/qwen-3-8-flash-next-vs-qwen-3-8-27b). So even if the weight file fits, inference may not. ### [Apple silicon](https://developer.apple.com/documentation/apple-silicon) and unified memory On Apple silicon, model weights share unified memory with macOS and the inference runtime[\[7\]](https://myclaw.ai/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b)[\[8\]](https://blog.laozhang.ai/ko/posts/qwen3-8-flash-next-vs-qwen3-8-27b). For more hardware options, see our [guide to local LLM models](/blog/best-local-llm-models-run). Fewer active parameters don’t guarantee faster performance when runtimes, context lengths, batch sizes, or quantization settings differ. Use that memory ceiling as the baseline when judging coding and agent performance in the next section. ## Coding, agents, and document tasks ### Coding scores and agent reliability Qwen’s published benchmarks put Flash Next ahead on agentic coding and tool use. Meanwhile, 27B stays close on single-turn coding and reasoning[\[2\]](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)[\[3\]](https://huggingface.co/Qwen/Qwen3.8-27B). | Benchmark | Task measured | Flash Next | 27B | | --- | --- | --- | --- | | [LiveCodeBench](https://livecodebench.github.io/) v6 | Code generation | 91.9 | 90.3 | | [DeepSWE 1.1](https://deepswe.datacurve.ai/) | Agentic coding | 58.7 | 42.2 | | [Toolathlon Verified](https://github.com/hkust-nlp/Toolathlon) | Tool use | 73.5 | 67.1 | | [GPQA Diamond](https://evalscope.readthedocs.io/en/v1.8.1/benchmarks/gpqa_diamond.html) | Reasoning | 91.7 | 89.2 | **Choose Flash Next** for jobs that span multiple tool calls, repo-wide edits, or multi-step agent loops. **Choose 27B** for shorter requests, local runs, or tighter VRAM budgets. These scores don’t guarantee agent reliability. Test both models with the same tool schema, retry limits, and output budget. That tradeoff matters most for coding agents; document-heavy tasks call for a different test. ### Long context and document RAG Both models support long-context work, but Flash Next handles extreme lengths more efficiently[\[2\]](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)[\[3\]](https://huggingface.co/Qwen/Qwen3.8-27B). Qwen reports that its Sparse Attention cuts prefill latency by **7.6x** and decode latency by **4.9x** at 1 million tokens[\[2\]](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)[\[9\]](https://arxiv.org/html/2608.30320). That makes Flash Next the better fit for long prompts and repo-scale RAG. The models are effectively tied on [OmniDocBench](https://github.com/opendatalab/OmniDocBench) 1.5[\[3\]](https://huggingface.co/Qwen/Qwen3.8-27B), so document understanding alone offers little reason to pick one over the other. For technical documentation and multi-step retrieval, test retrieval quality, chunking, and source citations on your own corpus before expanding the prompt. ### Hosted prices and GLM comparisons Hosted pricing also affects the choice. [Featherless](https://featherless.ai/) lists Qwen 3.8 Flash Next and [GLM-5.3-Flash](https://docs.z.ai/guides/vlm/glm-5.3-flash) at **$0.15 per million input tokens** and **$0.50 per million output tokens**. Qwen 3.8 27B costs $0.265 per million input tokens and $0.65 per million output tokens[\[1\]](https://featherless.ai/blog/glm-5-3-flash-vs-qwen-3-8-flash-next-vs-qwen-3-8-27b). GLM-5.3-Flash scores higher than Flash Next on DeepSWE 1.1 and Toolathlon Verified[\[1\]](https://featherless.ai/blog/glm-5-3-flash-vs-qwen-3-8-flash-next-vs-qwen-3-8-27b). Next, match the model to a local runtime and your available memory. ## Run each model with supported tools Once you’ve confirmed the model fits your hardware, choose a runtime that supports its weight format. ### Set up [Ollama](https://ollama.com/) and [LM Studio](https://lmstudio.ai/) For Qwen 3.8 27B, use this verified Ollama command:[\[5\]](https://www.respan.ai/articles/qwen-3.8-models-comparison)[\[10\]](https://aiprofitboardroom.com/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/) ```bash ollama run qwen3.8:27b ``` Flash Next also runs in Ollama. **Check the exact library entry** before pulling it.[\[10\]](https://aiprofitboardroom.com/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/) In LM Studio, search for Qwen 3.8. Before downloading, check the publisher, GGUF file, and runtime requirements.[\[10\]](https://aiprofitboardroom.com/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/) For setup steps, see our [local inference guide](/blog/running-llms-locally-ollama-llama-cpp-developers-guide). Need command-line control or server deployment? Use llama.cpp or vLLM instead. ### Set up [llama.cpp](https://github.com/ggml-org/llama.cpp) and [vLLM](https://vllm.ai/) llama.cpp supports both models. For 27B, use the GGUF file. Flash Next requires a build with **`qwen4_exp` support**.[\[6\]](https://kgptalkie.com/tutorials/llm-benchmarking/qwen-3-8-flash-next-vs-qwen-3-8-27b)[\[10\]](https://aiprofitboardroom.com/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/) ```bash llama-server -m qwen3.8-27b-q4_k_m.gguf ``` ```bash llama-server -m qwen3.8-flash-next-q4_k_xl.gguf --n-gpu-layers 48 ``` vLLM also supports both models:[\[2\]](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)[\[3\]](https://huggingface.co/Qwen/Qwen3.8-27B) ```bash vllm serve Qwen/Qwen3.8-27B ``` ```bash vllm serve Qwen/Qwen3.8-Flash-Next ``` ## Which model to choose by use case Based on the benchmarks and hardware tradeoffs above, use this table as a starting point. **Check the model card and your runtime limits before deploying.** | Use case | Recommended model | Reason | Prerequisite | | --- | --- | --- | --- | | Coding agents | Qwen 3.8 Flash Next | Best agentic performance. [\[2\]](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)[\[7\]](https://myclaw.ai/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b) | Hosted API or enough memory for the chosen checkpoint. | | Chat assistants | Qwen 3.8 27B | Simpler deployment and local privacy. [\[3\]](https://huggingface.co/Qwen/Qwen3.8-27B)[\[7\]](https://myclaw.ai/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b) | A supported quantization with room for chat history and runtime overhead. | | Document RAG | Qwen 3.8 Flash Next | Stronger long-context handling. [\[4\]](https://www.alibabacloud.com/blog/qwen-3-8-flash-next-a-new-architecture-towards-ultimate-cost-efficiency_603501)[\[11\]](https://atomic.chat/blog/guides/how-to-run-qwen-3-8-flash-next-locally) | Runtime support for the intended context length, plus YaRN scaling for very long prompts. | | On-device use | Qwen 3.8 27B | More likely to fit consumer hardware. [\[1\]](https://featherless.ai/blog/glm-5-3-flash-vs-qwen-3-8-flash-next-vs-qwen-3-8-27b)[\[7\]](https://myclaw.ai/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b) | 24 GB VRAM with a supported quantization. [\[1\]](https://featherless.ai/blog/glm-5-3-flash-vs-qwen-3-8-flash-next-vs-qwen-3-8-27b) | If your workload changes context length, quantization, or runtime, revisit the hardware sections above before committing. **Active parameters alone don't determine fit** - budget for weight size, context, and runtime headroom. ## FAQs ### Which Qwen 3.8 model is best for coding? For local coding, **choose Qwen3.8-27B** if you want a dense model that runs on typical consumer hardware. Choose **Qwen3.8-Flash-Next** if you have enough memory for a 125B mixture-of-experts model and want better coding and tool use results. Flash-Next leads on three published coding benchmarks: DeepSWE 1.1 (**58.7 vs. 42.2**), [SWE-bench Pro](https://www.swebench.com/) (**62.5 vs. 61.7**), and [NL2Repo-Bench](https://github.com/multimodal-art-projection/NL2RepoBench) (**48.1 vs. 42.3**) [\[1\]](https://featherless.ai/blog/glm-5-3-flash-vs-qwen-3-8-flash-next-vs-qwen-3-8-27b)[\[2\]](https://huggingface.co/Qwen/Qwen3.8-Flash-Next). ### Can Qwen 3.8 27B run on a 24 GB GPU? **Yes, Qwen 3.8 27B can run on a 24 GB GPU with Q4\_K\_M quantization.** This reduces its memory footprint from 55.6 GB to about 17 GB. That fits on a card like the [NVIDIA RTX 4090](https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/), though inference will be slower than with an uncompressed deployment. Watch your available GPU memory, especially during agentic tasks: memory use climbs as context length grows. ### What does Flash Next mean? Qwen3.8-Flash-Next is an open-weight mixture-of-experts model released as an early architectural preview [\[1\]](https://featherless.ai/blog/glm-5-3-flash-vs-qwen-3-8-flash-next-vs-qwen-3-8-27b)[\[2\]](https://huggingface.co/Qwen/Qwen3.8-Flash-Next). It serves as the base for the production-ready Qwen3.8-Flash API. The Flash-Next checkpoint gives developers a high-performance, cost-efficient model with **125 billion total parameters and 6 billion active parameters per token**. It’s optimized for long-context reasoning and agentic workflows [\[3\]](https://huggingface.co/Qwen/Qwen3.8-27B)[\[2\]](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)[\[4\]](https://www.alibabacloud.com/blog/qwen-3-8-flash-next-a-new-architecture-towards-ultimate-cost-efficiency_603501). ### Is Qwen 3.8 open weight and what is the licence? Both models are **open weight**, so you can download and inspect their weights. Their licenses differ: Qwen3.8-27B uses the permissive Apache 2.0 license, while Qwen3.8-Flash-Next uses the custom qwen-community-1.0 license. Before using Qwen3.8-Flash-Next in commercial products, **review the qwen-community-1.0 terms**. ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/og-image.png?v=a830cdf1","width":1200,"height":630},"sameAs":["https://twitter.com/dailydotdev","https://www.linkedin.com/company/dailydotdev","https://github.com/dailydotdev","https://www.instagram.com/dailydotdev"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","description":"Free, personalized developer news aggregator. Stay on top of software development news, AI coding tools, and web dev - curated daily from trusted sources.","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}},{"@type":"WebPage","@id":"https://daily.dev/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/","url":"https://daily.dev/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/","name":"Qwen 3.8 Flash Next vs Qwen 3.8 27B, which one to run and when | daily.dev","description":"Choose 27B for local/private runs and Flash Next for hosted coding agents or extreme long-context tasks; plan for VRAM, quantization, and runtime.","inLanguage":"en-US","isPartOf":{"@id":"https://daily.dev/#website"},"timeRequired":"PT6M"},{"@type":"Article","@id":"https://daily.dev/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/#article","headline":"Qwen 3.8 Flash Next vs Qwen 3.8 27B, which one to run and when","url":"https://daily.dev/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/","datePublished":"2026-11-03","dateModified":"2026-11-03T06:16:16.589Z","isPartOf":{"@id":"https://daily.dev/#website"},"publisher":{"@id":"https://daily.dev/#organization"},"mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/"},"description":"Choose 27B for local/private runs and Flash Next for hosted coding agents or extreme long-context tasks; plan for VRAM, quantization, and runtime.","image":{"@type":"ImageObject","url":"https://media.daily.dev/image/upload/s--oPUMc6tK--/f_auto,q_auto/v1/recruiter-landing/6ae96e8df0eee09effb2869c_1793684818134_65e92b395f?_a=BAMAMiB80"},"author":{"@type":"Person","name":"Kevin Nguyen"},"timeRequired":"PT6M","potentialAction":{"@type":"ReadAction","target":"https://daily.dev/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/"}},{"@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https://daily.dev/blog/"},{"@type":"ListItem","position":3,"name":"Trends","item":"https://daily.dev/categories/trends/"},{"@type":"ListItem","position":4,"name":"Qwen 3.8 Flash Next vs Qwen 3.8 27B, which one to run and when","item":"https://daily.dev/blog/qwen-3-8-flash-next-vs-qwen-3-8-27b/"}]},{"@type":"FAQPage","@context":"https://schema.org","mainEntity":[{"name":"Which Qwen 3.8 model is best for coding?","@type":"Question","acceptedAnswer":{"text":"For local coding, choose Qwen3.8-27B if you want a dense model that runs on typical consumer hardware. Choose Qwen3.8-Flash-Next if you have enough memory for a 125B mixture-of-experts model and want better coding and tool use results. Flash-Next leads on three published coding benchmarks: DeepSWE 1.1 (58.7 vs. 42.2), SWE-bench Pro (62.5 vs. 61.7), and NL2Repo-Bench (48.1 vs. 42.3).","@type":"Answer"}},{"name":"Can Qwen 3.8 27B run on a 24 GB GPU?","@type":"Question","acceptedAnswer":{"text":"Yes, Qwen 3.8 27B can run on a 24 GB GPU with Q4_K_M quantization. This reduces its memory footprint from 55.6 GB to about 17 GB. That fits on a card like the NVIDIA RTX 4090, though inference will be slower than with an uncompressed deployment. Watch your available GPU memory, especially during agentic tasks: memory use climbs as context length grows.","@type":"Answer"}},{"name":"What does Flash Next mean?","@type":"Question","acceptedAnswer":{"text":"Qwen3.8-Flash-Next is an open-weight mixture-of-experts model released as an early architectural preview. It serves as the base for the production-ready Qwen3.8-Flash API. The Flash-Next checkpoint gives developers a high-performance, cost-efficient model with 125 billion total parameters and 6 billion active parameters per token. It’s optimized for long-context reasoning and agentic workflows.","@type":"Answer"}},{"name":"Is Qwen 3.8 open weight and what is the licence?","@type":"Question","acceptedAnswer":{"text":"Both models are open weight, so you can download and inspect their weights. Their licenses differ: Qwen3.8-27B uses the permissive Apache 2.0 license, while Qwen3.8-Flash-Next uses the custom qwen-community-1.0 license. Before using Qwen3.8-Flash-Next in commercial products, review the qwen-community-1.0 terms.","@type":"Answer"}}]}]} ```