14 min read
Qwen3.8-27B
Alibaba27BOpen weights · August 14, 2026
Twenty-seven point seven eight billion parameters, an Apache 2.0 license, image and video input, and a 262,144-token window. The useful question is not whether it caught up with the closed models — that depends entirely on which one and which week — but what the last remaining reason was to keep the model out of your own machine, and whether that reason still holds. What Qwen measured, what an outsider measured, and what nobody has measured are kept apart here.
The thesis
On August 14, 2026 Alibaba shipped a dense 27B model under Apache 2.0 that scores 52 on the Artificial Analysis index: thirteen points above the closed flagship of the previous generation, six points below the one its own company had been selling for eleven days.
Spec sheet
- Shipped
August 14, 2026. The model card dates only the month; the day comes from the repository commit log, with the LICENSE landing at 04:15 UTC.
Vendor's own numberHugging Face · Qwen/Qwen3.8-27B · commits
- License
Apache 2.0. No active-user clause, no revenue threshold, no revenue share: the file in the repository is the full 2004 Apache text.
Vendor's own numberQwen/Qwen3.8-27B · LICENSE
- Parameters
27,781,427,952 in BF16. The card rounds that to “27B” and never reconciles the two figures. It is dense, so every one of them fires on every token.
Vendor's own numberHugging Face · Qwen/Qwen3.8-27B · safetensors
- Architecture
Causal Language Model with a vision encoder. 64 layers, hidden dimension 5120, and the hybrid block 16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN)).
Vendor's own numberQwen/Qwen3.8-27B · model card
- Vision encoder
The card publishes neither its size nor its architecture. Ollama weighs the projector separately — 461M parameters, 931MB — and that gap explains the two parameter counts in circulation. Qwen confirms no such split.
UnverifiedOllama · qwen3.8:27b
- Modalities
Text, image, and video in. Text only out.
Vendor's own numberQwen/Qwen3.8-27B · model card
- Context window
262,144 tokens natively, up to 1,000,000 with static YaRN, which the card itself warns may degrade shorter texts. The million-by-default belongs to Qwen Cloud, a paid service that had still not opened as of this reading.
Vendor's own numberQwen/Qwen3.8-27B · model card
- Context out of the box
LM Studio opens at 8,192 tokens and the model burns through them thinking. You have to raise it by hand to 262,144 before it stops hitting the ceiling.
Independently verifiedSimon Willison · Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things
- Default reasoning
reasoning_effort set to xhigh. The other two levels are medium and low, and thinking mode can be switched off per request.
Vendor's own numberQwen/Qwen3.8-27B · model card
- Size on disk
17.77GB at Q4_K_M, 29.12GB at Q8_0, 54.66GB at BF16. The quantization repository ships MTP layers and a vision projector alongside them.
Independently verifiedbartowski/Qwen3.8-27B-GGUF
- Memory in use
18GB of VRAM at 4k of context, 22GB at 64k, 34GB at the native 262,144 — on a 16.68GiB quant with vision switched off. The deployment costs twice what the file does.
Independently verifiedHardware Corner · We Tested Qwen3.8 27B
- Measured speed
40.31 tok/s on an RTX 3090 and 74.83 on an RTX 5090, both at 4k of context and fully offloaded to the GPU. The same 5090 drops to 22.79 tok/s at 128k.
Independently verifiedHardware Corner · We Tested Qwen3.8 27B
- Knowledge cutoff
Undeclared. Neither the model card nor any other primary source gives a training data cutoff. Printing one here would mean inventing it.
UnverifiedQwen/Qwen3.8-27B · model card
Whose work this changes
Software developer
Pull it, set reasoning_effort to medium before your first prompt, and time the answer rather than counting tokens per second.
Changes a lot
Data analyst
Run your own tables and charts through it before trusting the 83.7 on CharXiv: no one outside Qwen has measured this model’s vision yet.
Changes a little
Government office worker
Budget for 32GB of VRAM before you promise anything: the file weighs 18GB, but at full context the running model takes 34.
Changes a little
Reporter
Print the reasoning level next to every score: the same model reads 52, 44, or 43 depending on how it was started.
Changes a little
Teacher
Try it on the machine already in the room before requesting another one: under 32GB of unified memory it never reaches the window it advertises.
Changes a little
Lawyer
Nothing to do
Changes nothing
Graphic designer
Nothing to do
Changes nothing
Small shop owner
Nothing to do
Changes nothing
The shift is not the one being announced
The launch story is that an open model caught up with the closed ones. That story has been coming true in installments for about a year: the quality of what runs on hardware you own has been climbing slowly and without headlines since well before this release. A slope is not news. Something else happened on August 14, and it is the more interesting thing: every other reason to prefer the cloud ran out at once.
One reason is left standing, and it is the odd one out: speed. The other three were properties of a model, which is why a release could fix them. Speed is a property of the memory bandwidth of the machine underneath, and no checkpoint moves that. These are two different curves — one advances every few months in a repository, the other advances when somebody buys a different computer.
The numbers
The Artificial Analysis general index. The chart takes the benchmark leader and the three immediate neighbours above and below, so you see who it beats as well as who beats it. Every full name is in the table underneath.
63
58
56
53
52
44
39
OPUS 5
3.8 MAX
GEMINI 3.7
DEEPSEEK V4
3.8-27B · xhigh
3.8-27B · medium
3.7-PLUS
- Claude Opus 5 (max)63
- Qwen3.8 Max58
- Gemini 3.7 Flash (high)56
- DeepSeek V4 Pro53
- Qwen3.8-27B · xhigh52
- Qwen3.8-27B · medium44
- Qwen3.7-Plus39
This is the cleanest comparison in the dossier: every number comes from one method and one place. Eleven points to the frontier above. And thirteen points over the closed flagship this same house was selling in May, pinned in by hand because its job is not the distance but the jump.
Open weights
Independently verifiedArtificial Analysis · Intelligence Index
Speed is not tokens per second
Artificial Analysis measures 53.84 tok/s and flags it as notably slow against a class median near 105 — and it measures that against Alibaba’s own hosted API, not a local run. Hardware Corner measures 74.83 tok/s on a consumer RTX 5090 at 4k of context. A gaming card beats the vendor’s cloud, which is enough on its own to bury the idea that local is by definition the slow option.
And the rate itself, the part that really is a rate, is a hardware question. The model is dense, so all 27.78 billion parameters get read for every single token. That is where 40.31 tok/s on an RTX 3090 comes from, and 74.83 on a 5090, and 22.79 on that same 5090 once the context climbs to 128k. Speculative decoding through MTP adds somewhere between 33 and 181 percent depending on the card — another way of saying none of these numbers belong to the model.
What the announcement claims
- Plausible
“Context Length: 262,144 natively and extensible up to 1,000,000 tokens.”
True in config.json and false on startup. LM Studio opens at 8,192 tokens, Ollama makes you set the variable by hand, the million comes from static YaRN — which the card itself warns can hurt short inputs — and the million-by-default belongs to Qwen Cloud, still marked coming soon. Of the three numbers in circulation, the one you get without paying or configuring anything is the smallest.
Vendor's own numberQwen/Qwen3.8-27B · model card
- Marketing
“Matching the performance of Qwen3.7-plus — an MoE model ten times its size.”
That sentence lives in the press release, not in the technical documentation. And the size of Qwen3.7-Plus was never disclosed — Artificial Analysis notes that Alibaba has not published its parameter count — so “ten times” measures against a figure that exists nowhere. In the vendor’s own table, meanwhile, the 27B loses to that same Qwen3.7-Plus on HLE, GPQA Diamond, OmniDocBench 1.5, RealWorldQA, and ERQA.
Vendor's own numberAlibaba Cloud · Alibaba Unveils Qwen3.8-27B
- Plausible
“Problematic tasks were corrected, and all baseline models were re-evaluated on the refined benchmark.”
This is the SWE-bench Pro footnote, and it rewards a second reading: Qwen fixed the benchmark, then re-scored its own competitors on the version Qwen fixed. It did the same to the ground-truth annotations in MathVision and CharXiv. Saying so beats hiding it, and it still leaves all twenty-four rows without a single outside check. There is no technical report either: arXiv returns zero entries for Qwen3.8.
Vendor's own numberQwen/Qwen3.8-27B · model card
- Marketing
“xhigh (default): for complex tasks demanding thorough analysis.”
The setting shipped from the factory is the most expensive one available, and this dossier measures it further down. A default is not a specification: it is the vendor betting on what the average user wants, and it is the first bet worth undoing. Download this and leave it alone and you are not measuring the model, you are measuring the bet.
Vendor's own numberQwen/Qwen3.8-27B · model card
- Plausible
“A native vision-language model that understands images and videos.”
Verified as a statement, never as behavior. Not one outside measurement in this dossier touches an image or a video: Hardware Corner disabled vision for its runs, and Willison and Tunguz exercised text only. Locally the vision is not included either — the projector loads separately, 931MB that are not inside the 17GB file everyone keeps citing.
Vendor's own numberQwen/Qwen3.8-27B · model card
- Verified
“license: apache-2.0”
This is the one launch claim anybody can check in ten seconds without running a benchmark: the repository LICENSE is the complete Apache 2.0 text, and ModelScope registers the same license through the second official channel. It is also the claim that genuinely changes something — no user clause, no revenue threshold, no revenue share — and the one no headline bothers to carry.
Vendor's own numberQwen/Qwen3.8-27B · LICENSE
Leave the defaults alone and you will conclude local AI is bad
The GGUF preserves the factory xhigh, and that has visible consequences. Simon Willison asked it for an SVG of a pelican riding a bicycle: twenty-one minutes and 22,276 reasoning tokens to produce 3,223 tokens of output. The same request with reasoning turned off took 137 seconds. His recommendation is literal and leaves no room for a generous reading: ignore that default, start at low or at no reasoning at all.
A second default compounds the first. LM Studio opens with an 8,192-token window and the model eats it deliberating over trivia; Ollama needs somebody to set OLLAMA_CONTEXT_LENGTH by hand. Two settings from two different programs which, combined with the vendor’s dial, turn a good model into a bad demo. Anyone who downloads it this afternoon and changes nothing will not be measuring the model. They will be measuring three defaults.
What thinking harder costs
- Dial at xhigh, the factory setting
- 52Intelligence Index
- 160 million output tokens to run the whole index.
- Dial at medium, the sensible one
- 44Intelligence Index
- 75 million output tokens to run that same index.
Eight index points cost more than twice the tokens. The setting that ships from the factory is the most expensive one available, and it is the one anybody who downloads this today and changes nothing will end up measuring.
Independently verifiedArtificial Analysis · Intelligence Index
Where the 52 comes from, and what it does not mean
Nor was it measured on the weights you download. The page declares its performance source as Alibaba’s hosted API and never states which precision ran, while both a BF16 and an FP8 repository exist. Nobody has published an index for the 4-bit file. The object scoring 52 and the object sitting on a home SSD have not been shown to be the same thing, and that distinction is not pedantry — it is exactly what the headline rests on.
The comparisons need footnotes of their own. 52 against the 47 of Qwen3.7-Max is a win over an already retired flagship; against Qwen3.8-Max — the closed model Alibaba had been selling for eleven days when the 27B shipped — it reads 52 against 58. In blind human voting the gap to that flagship sits around 58 Elo points. And tying GPT-5.6 Luna means tying OpenAI’s cheap tier at a dollar per million input tokens, not the one they sell for hard problems.
Benchmark by benchmark
Terminal-Bench 2.1 · agent work in a terminal
88.2
82.9
82.7*
75.1
73
67.8
65.1
57
GLM-5.3
MUSE SPARK
DS V4 FLASH
DOTS3
3.8-27B
ORNITH 35B
POKEE 28B
LING 3.0
- GLM-5.3 (max)88.2
- Muse Spark 1.282.9
- DeepSeek V4 Flash 073182.7*
- dots3-note Preview75.1
- Qwen3.8-27B73
- Ornith-1.5-35B-A3B67.8
- Pokee-Isaac 28B65.1
- Ling 3.0 Flash57
* The vendor's own figure, measured on a different harness than the rest of the field.
Here the field actually opens up, 88 down to 46, and the 27B lands mid-table. That position is what makes a comparison useful: it loses to the ones above and beats the ones below, and you read both at once. The DeepSeek rows carry a mark because their own house publishes them on a different harness, as the leaderboard warns.
Open weights
Independently verifiedTerminal-Bench 2.1 · leaderboard
GPQA Diamond · graduate-level science questions
95.5
92.9
92.4
89.9
88.9
88.1
78.5
68.4
47
GEMINI 3.1
KIMI K3
DEEPSEEK V4
DS V4 FLASH
3.8-27B
GLM-5.3
GPT-OSS
MISTRAL L
LLAMA 4
- Gemini 3.1 Pro Preview95.5
- Kimi K392.9
- DeepSeek V4 Pro92.4
- DeepSeek V4 Flash 073189.9
- Qwen3.8-27B88.9
- GLM-5.3 (max)88.1
- gpt-oss-120b (high)78.5
- Mistral Large 251268.4
- Llama 4 Scout47
The shape of this chart IS the argument, and it has to be read against the one above. On exam questions the field is packed tight; on terminal work it runs from 46 to 88. Knowing is not doing, and doing is the part you pay for. Llama 4 Scout is pinned in, far below everyone, because without the floor of the open catalogue a field of neighbours lets you believe everything open sits around ninety. And here a third party measured the 27B, so for once its number carries no asterisk.
Open weights
Independently verifiedVals AI · GPQA Diamond (espejo de BenchLM, 19 ago 2026)
What it costs to keep at home
“It fits in 17GB” describes a file at rest, not a deployment. Ollama pulls 18GB plus 931MB of vision projector, and memory in use grows with the window: 18GB at 4k, 22 at 64k, 26 at 128k, and 34GB at the native 262,144 with vision off. This model in particular cannot be run at a short context, because a short context is what it spends on thinking. The configuration that produces the quality wants twice the memory of the headline.
All that said: 2,358,347 downloads of the official repository nine days after launch, with the Unsloth GGUF mirror counting 6,674,515 on its own. Caveats intact, a great many people already have the file. That is the part of the story that needs no benchmark to be true.
The verdict
Qwen3.8-27B does not prove that open weights caught up with closed ones. It proves that the last reason to prefer the cloud stopped being a model and became a graphics card.