by Alibaba Cloud · Qwen3.8 family · best for strongest Apache-2.0 model that fits one GPU
Qwen3.8-27B is the dense, single-GPU member of Alibaba Cloud's Qwen3.8 family, published with open weights under Apache 2.0 on August 14, 2026 and served on Model Studio from August 17. It is a native vision-language model (text, image and video in), thinks by default with adjustable effort, runs 262,144 tokens natively and 1M tokens hosted, and succeeds the Qwen3.6-27B and the older Qwen3-32B. Independent testing backs the headline: Artificial Analysis ranks it first of 142 open-weight models in the 4B-40B size class, with an Intelligence Index of 34, roughly double the scores of Gemma 4 31B, Muse Glimmer 30B, Nemotron 3.5 Lightning and Mistral Small 4. For a buyer: this is the model to self-host when you want near-frontier reasoning and coding on one GPU with no license strings. Alibaba's own hosted version ($0.50 / $3.00 per 1M tokens) is slow and costs more than the stronger Qwen3.8-Flash.
reasoning_effort (xhigh, medium, low) and preserve_thinking to keep reasoning across agent turns.| Benchmark | Score | Source |
|---|---|---|
| Humanity's Last Exam | 33.9% | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
| LMArena Elo | 1437 | lmarena.ai 2026-09-13T00:00:00.000Z |
| GPQA Diamond | 90.5% | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
| LiveCodeBench | 90.3% | huggingface.co 2026-08-14T00:00:00.000Z |
| Terminal-Bench | 79.8% | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
| LMArena Coding Elo | 1499 | lmarena.ai 2026-09-13T00:00:00.000Z |
| Artificial Analysis Index | 34 | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“The default open model for a single GPU: Apache 2.0, first in its class on independent tests, and no license review needed before you ship.”
Qwen3.8-27B is an easy strategic choice for anyone standardizing on open weights. It leads its size class by a wide margin on Artificial Analysis, it is licensed Apache 2.0 with no user or revenue thresholds, and it fits on one GPU, so the vendor-risk question largely disappears once the weights are downloaded. The Qwen line's release cadence (Qwen3.5 in February to Qwen3.8 in August) gives good roadmap confidence, though it also means frequent re-evaluation. The one caution is to avoid Alibaba's hosted version as a production default: it is slow, and Qwen3.8-Flash is better value on that platform.
“Alibaba keeps its most permissive license for the model the community uses most, and at this size nobody else is close on the independent index.”
The 27B dense size is where open-model demand concentrates: large enough for serious work, small enough for one card. Qwen owns it. Artificial Analysis puts it at 34 against 19 for Gemma 4 31B and 17 for Muse Glimmer 30B, Meta's return to open weights, which Alibaba's own model card already benchmarks against. Keeping this tier on Apache 2.0 while moving Flash-Next and Max weights to custom licenses is a deliberate split: maximum reach for the community model, protection for the commercial tiers. The open flank is human preference, where Gemma still wins on LMArena text.
“Free to self-host on one GPU and pricey to rent from Alibaba. The economics depend almost entirely on whether you run the weights yourself.”
As weights, the cost is hardware only: one 80 GB GPU at BF16 or a 40-48 GB card at FP8, with no license fee or usage cap. Through Alibaba, the numbers are poor. $0.50 / $3.00 per 1M tokens is three to six times the class median of $0.16 / $0.53, and heavy reasoning pushed Artificial Analysis's cost to $0.82 per index task against $0.37 for Qwen3.8-Flash, which scores higher. Third-party hosts set their own prices, which you need to quote before committing. Budget for verbosity either way: at xhigh effort it produced 200M output tokens where the class median is 82M.
“It loads in vLLM or SGLang on one card, reads images and video, and has effort control. Leave the default context alone unless you need 1M.”
The operating experience is good. BF16 and official FP8 checkpoints, recipes for SGLang, vLLM and TokenSpeed, and an OpenAI-compatible surface mean a production endpoint is quick to stand up. Thinking can be disabled per request, reasoning_effort trims depth, and preserved thinking helps multi-turn agents. Watch three things. The tool-call format is XML-style and new, so older Qwen3 parsers need updating. Going past 262K means static YaRN, which Alibaba says can dent short-prompt quality. And the vocabulary grew from 151,936 to 248,320 tokens, so existing Qwen3 fine-tunes and tokenizer-dependent tooling do not carry over.
“Smart and careful, with fewer confident wrong answers than most, but through Alibaba's API it is slow and wordy, with long pauses before answers.”
Answer quality punches well above 27B: strong science and coding reasoning, good screenshot and document reading, and a notably low hallucination rate on Artificial Analysis's knowledge test. Day to day, the hosted version tests patience: about 46 tokens/sec and roughly 47 seconds of thinking before the first answer token at default effort. Human raters on LMArena place it mid-pack at 1437, below Gemma 4 31B, which suggests its style is less engaging in chat than its scores imply. Run it locally on a fast GPU, or switch thinking off for simple questions, and the experience improves.
“A month-old model whose class lead is confirmed by one independent lab, while its best numbers come from Alibaba's own harness and in-house suites.”
For a model weeks old, the independent record is unusually solid. Artificial Analysis confirms the class lead and actually measures higher GPQA, HLE and Terminal-Bench 2.1 than Alibaba claims, and LMArena has over 10,000 votes. The weaker parts are Alibaba's. QwenSWEBench and CoWorkBench are in-house, SWE-bench Pro ran on a task set Alibaba corrected itself, and the comparisons use Alibaba's harness for every baseline. Real weaknesses show up independently: 5.6% on Terminal-Bench 4.0, heavy verbosity, and a human-preference rating below Gemma 4 31B. The class lead looks real, but benchmark leadership is not the same as users preferring it.
The full research notes behind this review — verified against primary sources.
A dense decoder with a vision encoder: 27.8B parameters in the published BF16 checkpoint, 64 layers and a hidden size of 5,120. The layers repeat a block of three Gated DeltaNet layers followed by one gated-attention layer, each with a feed-forward network (intermediate size 17,408). Gated DeltaNet layers use 48 value heads and 16 query/key heads of dimension 128. Full-attention layers use grouped-query attention with 24 query heads and 4 key-value heads of dimension 256, and rotary embeddings on 64 dimensions. Only 16 of the 64 layers keep a growing KV cache, which keeps long-context memory modest. A multi-token-prediction head supports speculative decoding. The vocabulary is 248,320 tokens (padded), up from Qwen3's 151,936. Native context is 262,144 tokens; Alibaba documents static YaRN with a factor of 4 to reach 1M and warns it can slightly reduce quality on short texts, so set the factor to the length you actually need. The Hugging Face model_type is qwen3_5, the same architecture family as Qwen3.5 and Qwen3.6. Training data, token counts and knowledge cutoff are not disclosed. Alibaba's published base-model scores for Qwen3.8-27B-Base include MMLU 87.5, MMLU-Pro 68.6 and BBH 89.6.
Coding (8.3): Alibaba reports SWE-bench Pro 61.7 (Muse Glimmer 30B: 51.2 on the same harness), Terminal-Bench 2.1 73.0 on the Terminus harness, LiveCodeBench v6 90.3 and its in-house QwenSWEBench 79.0. Artificial Analysis measures 79.8% on Terminal-Bench 2.1 and 46.6% on SciCode but only 5.6% on the much harder Terminal-Bench 4.0, where Qwen3.8-Flash reaches 25.3%. On LMArena's coding board it scores 1499, level with Gemma 4 31B. Reasoning (8.3): GPQA Diamond 90.5% and Humanity's Last Exam 33.9% in Artificial Analysis testing, above the vendor's 89.2 and 30.8. Math (8.0): MathVision 90.0 (94.6 with code interpreter); no AIME result is published. Agentic (8.0): vendor-reported OSWorld-Verified 84.3, WebArena-Verified 64.8, AndroidWorld 81.9 and CoWorkBench 70.7. Artificial Analysis records 48% on its agentic banking tool-use test, 48% on AutomationBench-AA and a GDPval-AA Elo of 1409. Long context (7.8): 82% on AA-LCR, slightly above Qwen3.8-Flash's 80%. That holds on the hosted 1M window, but self-hosters must configure YaRN beyond 262K. Multilingual (8.3): the base model scores MMMLU 79.7, MGSM 86.4 and INCLUDE 74.4 on Alibaba's table. Vision (8.3): MMMU-Pro 76.3% (Artificial Analysis), RealWorldQA 85.9 and CharXiv reasoning 83.7 (vendor). Document and OCR (8.5): OmniDocBench 1.5 91.1 against Muse Glimmer's 75.8 on Alibaba's table, plus a hosted PDF-parsing tool. Creative writing (7.0): the LMArena text rating (1437) sits below Gemma 4 31B's 1451. Instruction following (8.2): IFBench 79.5 (vendor). Function calling (8.2): XML-style tool calls with multiple calls per turn in the chat template, and structured outputs on the hosted API. Safety calibration (7.3): Artificial Analysis measures a 30% hallucination rate, the best among the small open models compared here. Realtime data (5.0): built-in web search on Model Studio only; the weights have none.
| Benchmark | Score | vs Predecessor | vs Top Competitor | Source |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v4.3 | 34 at xhigh (28 medium, 26 low, 20 non-reasoning) | Qwen3-32B reasoning 9; Qwen3.6-27B 21 | #1 of 142 open models in its size class; Gemma 4 31B 19, Muse Glimmer 17, Nemotron 3.5 Lightning 13, Mistral Small 4 11 | Artificial Analysis |
| GPQA Diamond | 90.5% (Artificial Analysis); 89.2 (vendor) | Qwen3-32B 66.8%; Qwen3.6-27B 84.2% | Gemma 4 31B 85.7%, Muse Glimmer 83.5%, Mistral Small 4 76.9% | Artificial Analysis |
| Humanity's Last Exam | 33.9% (Artificial Analysis); 30.8 (vendor) | Qwen3-32B 7.4%; Qwen3.6-27B 23.1% | Gemma 4 31B 23.6%, Muse Glimmer 22.0%, Nemotron 3.5 Lightning 10.6% | Artificial Analysis |
| Terminal-Bench 2.1 | 79.8% (Artificial Analysis); 73.0 (vendor, Terminus) | Qwen3.6-27B 60.7% | Muse Glimmer 51.7%, Gemma 4 31B 43.4%, Nemotron 3.5 Lightning 24.3% | Artificial Analysis |
| Terminal-Bench 4.0 | 5.6% | Qwen3.6-27B 0% | every peer here scores 1% or less; Qwen3.8-Flash 25.3% | Artificial Analysis |
| MMMU-Pro | 76.3% | Qwen3.6-27B 74.6% | Muse Glimmer 74.3%, Gemma 4 31B 73.4%, Mistral Small 4 56.8% | Artificial Analysis |
| AA-LCR (long context) | 82.0% | Qwen3.6-27B 77.3% | Muse Glimmer 83.3%, Gemma 4 31B 69.7% | Artificial Analysis |
| LMArena Text | 1437 ±6 (10,697 votes, #90 of 402) | Qwen3-32B 1347 ±9 | Gemma 4 31B 1451 ±8, Muse Glimmer 1427 ±10 | LMArena |
| LMArena Coding | 1499 ±11 | not listed | Gemma 4 31B 1499 ±15, Muse Glimmer 1483 ±19 | LMArena |
| LiveCodeBench v6 | 90.3 (vendor) | Qwen3.6-27B 83.9 | Claude Opus 4.6 (Max) 88.8 on Alibaba's table | Hugging Face card |
| SWE-bench Pro | 61.7 (vendor) | Qwen3.6-27B 53.5 | Muse Glimmer 30B 51.2 on the same harness | Hugging Face card |
| OSWorld-Verified | 84.3 (vendor) | Qwen3.6-27B 63.9 | Muse Glimmer 30B 65.9 | Hugging Face card |
| OmniDocBench 1.5 | 91.1 (vendor) | Qwen3.6-27B 89.4 | Muse Glimmer 30B 75.8 | Hugging Face card |
| IFBench | 79.5 (vendor) | Qwen3.6-27B 69.1 | Muse Glimmer 30B 77.0 | Hugging Face card |
Artificial Analysis measured the model through Alibaba's API at the default xhigh effort. Alibaba ran its own table on the Claude Code harness at temperature 1.0 and a 256K window, over a SWE-bench Pro task set it corrected and re-ran for all baselines. It judged HLE with GPT-4o, and QwenSWEBench and CoWorkBench are in-house. The Artificial Analysis index v4.3 is a new scale and is not comparable with older index figures in this catalog. MMLU, MMLU-Pro and BBH exist only for the base model, and no SWE-bench Verified, AIME, MMMU or IFEval figure is published, so those fields are null.
Through Alibaba's API, Artificial Analysis measures 46.4 output tokens/sec and 3.94 seconds to the first token, against class medians of 89.2 tokens/sec and 2.12 seconds. At the default xhigh effort the first answer token arrives after about 47 seconds. Dropping to medium or low effort only shortens that to about 42 seconds, and non-thinking mode runs at 49 tokens/sec with a 3.9-second first token. It is also verbose: 200M output tokens to run the Intelligence Index, against a class median of 82M. Hosted rivals are much quicker in the same testing: Nemotron 3.5 Lightning at about 297 tokens/sec, Mistral Small 4 at 155 and Muse Glimmer at 91. Self-hosted speed depends on your hardware; the multi-token-prediction head and the linear-attention layers help, but no independent self-hosted measurement exists yet.
| Surface | Cost | Notes |
|---|---|---|
| API input | $0.50 / 1M tok | Singapore (International), flat 0 to 1M, thinking and non-thinking |
| API output | $3.00 / 1M tok | Chain-of-thought tokens bill as output |
| Cached input | $0.05 / 1M tok | Explicit cache read; implicit cache hits bill $0.10 |
| Cache write | $0.625 / 1M tok | Explicit cache creation |
| Batch (in/out) | not offered | Batch Inference is listed as unsupported |
| China (Beijing) | $0.424 / $1.696 | Mainland region list price |
| Free tier | 1M tokens | Singapore only, valid 90 days |
| Rate limits | 5K RPM / 5M TPM | Qwen Cloud listing; Singapore and Beijing use dynamic limits tiered by monthly spend |
| Self-host | no license fee | Apache 2.0; one 80 GB GPU at BF16 or a 40-48 GB GPU at FP8 |
| Third-party hosts | set by each host | DeepInfra, Novita and OVHcloud are live through Hugging Face's inference-provider routing |
Artificial Analysis calls Alibaba's rate expensive for the size class, where the median is $0.16 input / $0.53 output. Running its index cost $1,170, or about $0.82 per task, more than twice Qwen3.8-Flash's $0.37 per task. Qwen3.8-Flash also scores higher (40 vs 34) at $0.15 / $0.47. On Alibaba's own cloud, the 27B is rarely the economical API choice. Its economics come from running the weights yourself or through a third-party host.
Self-hosting is the main route. The Apache 2.0 weights are on Hugging Face as BF16 (about 56 GB) and an official FP8 build (about 31 GB). BF16 fits one 80 GB GPU, and FP8 fits a single 40-48 GB card with room for cache. Alibaba provides recipes for SGLang, vLLM and TokenSpeed, and the model loads in Hugging Face Transformers. Alibaba does not publish 4-bit builds. For hour-scale video, Alibaba recommends raising longest_edge in the video preprocessor config to allow about 224K video tokens. Beyond 262,144 tokens you enable YaRN through rope_parameters or engine flags. Through Hugging Face's inference-provider routing, DeepInfra, Novita and OVHcloud serve it live, and Artificial Analysis counts eight API providers in total. Managed first-party access is Alibaba Cloud Model Studio and Qwen Cloud (Beijing and Singapore regions), with OpenAI-compatible and DashScope APIs, function calling, structured outputs, context caching and built-in web search, code interpreter and PDF parsing. Qwen Cloud labels it "Qwen3.8-Open-Source" with the ID qwen3.8-27b.
Apache 2.0 means no usage restrictions, monthly-user thresholds or naming obligations beyond keeping the license and notices. That is a cleaner position than Qwen3.8-Flash-Next's Qwen Community License. Alibaba publishes no safety framework, system card or refusal data for the model, and no compliance certifications on its model pages. Whether hosted API inputs are used for training is not stated there. Self-hosting removes data egress entirely. The hosted endpoint processes requests in Singapore or mainland China (Beijing). Artificial Analysis's AA-Omniscience test gives the one independent calibration signal: a 30% hallucination rate. That is the lowest among the small open models compared here (Nemotron 3.5 Lightning 38%, Mistral Small 4 67%, Muse Glimmer 82%, Gemma 4 31B 85%), though overall knowledge accuracy is modest at 16%.
Hugging Face lists about 7.2M downloads of the BF16 weights and 6.6M of the FP8 build in the past month, with 16K likes, far ahead of any other Qwen3.8 release. Alibaba provides serving recipes for SGLang, vLLM and TokenSpeed, and the model loads in Hugging Face Transformers. DeepInfra, Novita and OVHcloud serve it through Hugging Face's inference-provider routing, and Artificial Analysis tracks eight API providers. First-party access is through Alibaba Cloud Model Studio and Qwen Cloud, with Python and Node.js examples for the OpenAI SDK, the DashScope SDK and cURL. LMArena rates it with more than 10,000 votes, which says it is widely used in practice.
The weights are free under Apache 2.0. Alibaba's hosted API charges $0.50 input and $3.00 output per 1M tokens in Singapore, with cache reads at $0.05 (explicit) or $0.10 (implicit) and 1M free tokens for 90 days. The Beijing list price is $0.424 / $1.696, and DeepInfra, Novita and OVHcloud set their own rates.
The BF16 checkpoint is about 56 GB and fits one 80 GB GPU. The official FP8 build is about 31 GB and fits a single 40-48 GB GPU with room for cache. Alibaba publishes no 4-bit build.
Yes. The LICENSE file in the Hugging Face repository is the standard Apache License 2.0, copyright Alibaba Cloud, with no user-count or revenue clauses. That differs from Qwen3.8-Flash-Next, whose Qwen Community License restricts model-hosting and coding/office-assistant businesses.
On Alibaba's API, Qwen3.8-Flash is cheaper ($0.15 / $0.47), scores higher on the Artificial Analysis Index (40 vs 34) and is slightly faster. Choose the 27B when you need Apache 2.0 weights, single-GPU self-hosting or fine-tuning.
1M tokens on the hosted API. The weights run 262,144 tokens natively and reach 1M with YaRN, which Alibaba advises enabling only when needed because it can slightly reduce short-text quality.
Yes, natively, including documents, charts and hour-scale video. For long video, raise longest_edge in the video preprocessor config as Alibaba recommends.
Disable thinking per request or lower reasoning_effort. Artificial Analysis scores it 20 in non-thinking mode and 34 at xhigh, so test which depth your task needs. Alibaba notes that low effort can lengthen multi-turn agent runs through retries.
Not directly. The architecture (hybrid Gated DeltaNet attention), tokenizer (248,320 vs 151,936 tokens) and tool-call format all changed, so plan to re-run fine-tuning and update tool parsers.
Meta's Apache 2.0 dense model (29.6B, text and image in, 128K context), released August 10, 2026, with a 4-bit build for 24 GB GPUs. Qwen leads on the Artificial Analysis Index (34 vs 17), GPQA (90.5% vs 83.5%), HLE (33.9% vs 22.0%) and hallucination rate (30% vs 82%), and on Alibaba's table (SWE-bench Pro 61.7 vs 51.2, OmniDocBench 91.1 vs 75.8). Glimmer is about twice as fast in hosted testing (91 vs 46 tokens/sec), slightly ahead on AA-LCR (83% vs 82%) and easier to fit on consumer hardware.
Apache 2.0 119B mixture-of-experts with a first-party API at $0.15 / $0.60, 256K context and image input. It is far cheaper to rent and faster (about 155 tokens/sec), but scores 11 vs 34 on the Artificial Analysis Index and 76.9% vs 90.5% on GPQA, and it needs about an 80 GB GPU at FP8 against the 27B's roughly 31 GB.
A 30B-total, 3B-active hybrid Mamba-2 MoE under the OpenMDW-1.1 license with up to 1M native context. It is built for speed (about 297 tokens/sec and a 0.6-second first token in Artificial Analysis testing) and costs little to serve. It is text-only and scores 13 vs 34 on the index, with a preliminary LMArena rating of 1351.
Google's dense 31B open model from March 31, 2026, with up to 256K context and open weights that Google says permit responsible commercial use. Human raters prefer it on LMArena text (1451 vs 1437) and tie it on coding (1499), but it trails on the Artificial Analysis Index (19 vs 34), GPQA (85.7% vs 90.5%) and HLE (23.6% vs 33.9%), and hallucinates far more (85% vs 30%).
Primary references used to verify this review.
Last verified 2026-09-21