Qwen3.8-27B is a dense 27B vision-language model built for 2026's wave of single-GPU agentic tools, not a mixture-of-experts design like its 2.4T sibling. It scores 84.3% on the OSWorld-Verified computer-use benchmark and suits developers who want a self-hostable, open-weight coding agent instead of a closed API.
Qwen3.8-27B is a 27.78-billion-parameter dense vision-language model Alibaba released on August 14, 2026 with open weights. It scores 89.2% on GPQA Diamond, reading text, images, and video natively across a 262,144-token context window extendable to 1 million tokens via YaRN.
Provider: Alibaba Cloud (Qwen Team) · Family: Qwen
More about Alibaba Cloud (Qwen Team) on HokAI
Context window: 262,144 tokens · Max output: 131,072
Input modalities: text, image, video · Output: text
About Qwen3.8-27B
Qwen3.8-27B is a dense, natively multimodal large language model built by Alibaba Cloud's Qwen team (Tongyi Lab) and released on August 14, 2026 under the Apache 2.0 license. The checkpoint carries 27.78 billion parameters spread across 64 transformer layers, arranged in a hybrid pattern of 48 Gated DeltaNet linear-attention blocks and 16 full-attention blocks. Alibaba positions it as the self-hostable sibling of the Qwen3.8-2.4T-A95B mixture-of-experts flagship, for developers who want frontier-adjacent coding ability without a data center. On graduate-level science reasoning, Qwen3.8-27B scores 89.2% on GPQA Diamond, up from 87.8% on the prior Qwen3.6-27B generation. It posts 61.7% on SWE-bench Pro, 73.0% on Terminal-Bench 2.1, and 90.3% on LiveCodeBench v6, and reaches 30.8% on Humanity's Last Exam without tools. Alibaba's own comparison shows it ahead of Meta's Muse Glimmer 30B on every overlapping text benchmark it reports, including a 41% relative margin on Terminal-Bench 2.1 versus Glimmer's 51.7%. Closed frontier models like Claude Opus 4.6 still lead on the hardest evaluations, so Qwen3.8-27B competes at the top of the open-weight class rather than the absolute frontier. The model ships with a native 262,144-token context window, extendable to 1,048,576 tokens through YaRN, with a recommended 131,072-token cap on the final response for agentic workloads. Static YaRN can reduce quality on shorter prompts, so the extension suits genuinely long documents or video rather than everyday use. Qwen3.8-27B reads text, images, and video natively, covering STEM diagrams, scanned documents, and hour-scale video without a separate vision adapter. It scores 84.3% on OSWorld-Verified and 64.8% on WebArena-Verified, benchmarks that test whether a model can operate a desktop or browser interface from screenshots, and 94.6% on MathVision with a code interpreter. Output stays text only. A built-in multi-token-prediction head supports speculative decoding, and a thinking-mode toggle trades latency for accuracy on hard multi-step problems. Alibaba has not listed a first-party metered API for this checkpoint, unlike the Qwen3.8-Max sibling, priced at $2/$6 per million input/output tokens. Third-party marketplaces filled that gap within a day, pricing the 27B checkpoint well under the Max rate; exact current rates are in the pricing FAQ below. Because the weights are free, the realistic cost for most adopters is GPU rental rather than a per-token fee. The official BF16 weights total about 51.8GiB, needing roughly 68GiB to run the full 262K-token context on an 80GB-class GPU such as an H100. An FP8 build cuts that to about 28.8GiB, and the community shipped quantized builds within hours that fit a single 24GB to 48GB consumer card. Supported serving stacks include vLLM, SGLang, llama.cpp, and Transformers, and AMD published day-zero support notes for its Ryzen AI Max and Radeon GPU lines. Alibaba did not publish a system card, training cutoff, or safety evaluation for this checkpoint, unusual for a release this size. Within a day, community fine-tunes with reduced refusal behavior appeared on Hugging Face, typical where end users control moderation rather than a vendor API. Qwen3.8-27B suits developers building a self-hosted coding agent, privacy-sensitive or air-gapped teams that cannot send data to a third-party API, and researchers needing a long-context multimodal model on one GPU. Enterprises that require a vendor SLA, a published system card, or scientific reasoning above the 90 percent mark are better served by a closed frontier model such as Claude Opus 4.6 or GPT-5.5. Qwen3.8-27B is the direct successor to Qwen3.6-27B and arrived alongside the larger Qwen3.8-2.4T-A95B model, the base checkpoint behind Qwen3.8-Max. Alibaba's Qwen team has shipped a major release roughly every one to two months through 2026, so a further point release is a reasonable expectation based on cadence, not any confirmed roadmap.
Pricing
Open weights are free to download and self-host under a permissive license, so the real cost is GPU time rather than a token fee. Alibaba has not published its own managed-API rate for this specific checkpoint; the single third-party host listed on OpenRouter currently prices it at $0.45 per 1M input tokens and $3.20 per 1M output tokens, well under Alibaba's own $2/$6 pricing for the larger Qwen3.8-Max.
Key Features
- Hybrid Linear/Full Attention: Three linear-attention layers run for every one full-attention layer across the transformer stack, cutting the memory cost of long-context inference versus a pure transformer design.
- 262K Native Context, 1M via YaRN: Ships with a native long context window that Alibaba documents as extendable roughly fourfold through YaRN scaling for long-document and long-video work.
- Native Vision and Video Input: Reads images, STEM diagrams, documents, and hour-scale video directly, without a separate vision adapter model.
- Flexible Thinking Control: Toggles visible reasoning on or off per request, trading latency for accuracy on hard agentic and math tasks.
- Fully Open, Permissive License: Weights are free for commercial use, modification, and redistribution under a permissive open-source license, with no gated approval process.
Pros
- Outpaces Meta's Muse Glimmer 30B across the board on Alibaba's published text benchmarks, by as much as 41 percentage points on Terminal-Bench 2.1.
- Improves on its own predecessor Qwen3.6-27B across every reported benchmark row, with double-digit point gains on agentic computer-use tests.
- A single high-end consumer GPU can run the FP8 build, whose 28.8GiB footprint is roughly half the size of the full-context memory need.
Cons
- Trails closed frontier models on the hardest reasoning evaluations, despite leading its own open-weight size class.
- No official Alibaba Cloud managed API or pricing was confirmed for this checkpoint at launch; only third-party hosts list per-token rates.
- No published system card: training data cutoff, safety evaluation, and supported-language count are undisclosed.
Benchmarks
- gpqa diamond: 89.2
- swe bench pro: 61.7
- livecodebench v6: 90.3
- osworld verified: 84.3
- webarena verified: 64.8
- mathvision with ci: 94.6
- terminal bench 2 1: 73
- humanitys last exam: 30.8
Frequently Asked Questions
How much does Qwen3.8-27B cost to run?
Qwen3.8-27B's weights are free under an open license, so self-hosting costs only GPU time. Through the third-party marketplace OpenRouter, the sole listed host charges around $0.45 for every 1M input tokens and $3.20 for every 1M output tokens as of August 2026, since no first-party metered endpoint exists yet for this checkpoint.
How does Qwen3.8-27B compare to other open-weight models?
Alibaba's own published comparison puts Qwen3.8-27B ahead of Meta's similarly sized Muse Glimmer model on every overlapping text benchmark it reports. Independent trackers note that closed frontier labs like Anthropic and OpenAI still lead on the hardest reasoning evaluations, so its edge is strongest within the open-weight class rather than against paid frontier APIs.
Is Qwen3.8-27B open source or proprietary?
Qwen3.8-27B is fully open-weight under the Apache 2.0 license, one of the most permissive licenses available, so commercial use, modification, and redistribution are all allowed. Weights are downloadable from Hugging Face in full precision and several community-quantized formats.
Does Qwen3.8-27B train on user data?
Alibaba has not published a system card or a managed-API data policy for this checkpoint, so there is no confirmed first-party retention policy to cite. Self-hosted deployments keep all data on the operator's own hardware, and third-party inference hosts set their own independent retention terms.
Who is Qwen3.8-27B best for, and who should avoid it?
It fits teams that want a self-hostable coding or computer-use agent without ongoing per-token fees, plus anyone needing native image or video understanding on their own hardware. Organizations that need a vendor-backed service agreement, documented safety testing, or the very top tier of graduate-level reasoning accuracy should pick a closed frontier model instead.