ExplainedQwen 3.8Qwenopen-weight modelsOllamalocal LLMsthinking models

Qwen 3.8 explained: what is new, and running the 27B model with Ollama

Qwen 3.8 is the newest family of AI models from Alibaba's Qwen team. Three of its models have open weights: files you can download and run yourself. I ran the smallest one, Qwen3.8-27B, with Ollama, a free tool that runs models on your own machine, and recorded every step. This page covers what is new, what it needs, and what I measured.

Roni Das

14 min read

Published6 October 2026, 12:10 UTC

Two courses by the author of this page:

AI Engineering

204 lessons · 10 free to read

₹999 in India · $49 elsewhere, once

Get AI Engineering

You own this course

System Design

770 lessons · 18 free to read

₹499 in India · $49 elsewhere, once

Get System Design

You own this course

Every lesson in both courses has an AI tutor beside it. It reads the same lesson you are reading and answers your questions from it. Try the tutor

Three Qwen 3.8 open-weight models drawn as stacks of different heights: Qwen3.8-27B, 27 billion parameters, Apache 2.0, 18 GB in Ollama; Qwen3.8-Flash-Next, 125 billion with 6 billion active; Qwen3.8-2.4T-A95B, 2.4 trillion with 95 billion active. A label marks the 27B as the one we ran. Stack heights show order of size, not scale. — Three open models, very different sizes

In short

Qwen 3.8 is a family, not one model. The open-weight models are Qwen3.8-27B, Qwen3.8-Flash-Next and Qwen3.8-2.4T-A95B; Qwen3.8-Max, Qwen3.8-Flash and Qwen3.8-Omni-Flash are services you call over the internet.
Only Qwen3.8-27B uses the Apache 2.0 licence, legal terms that allow use in paid products. The two big open models have their own Qwen licences, so read them before using them in a product.
The models think before they answer, unless you turn that off per request, and you can set how hard they think. The 27B model also reads images and videos.
The smallest Qwen 3.8 download in Ollama's library is qwen3.8:27b, about 18 GB. On a rented NVIDIA L4 GPU (graphics card) it wrote 18 to 35 tokens (small pieces of words) a second once loaded, depending on the text, in single runs. It read all 5 values from one simple bar chart.
On Qwen's own tests the 27B scores well above the older Qwen3.6-27B, for example 61.7 against 53.5 on SWE-bench Pro, a test of fixing real GitHub issues. Qwen ran these on its own setup, not on the 4-bit Ollama copy; I did not rerun them.

First model: Qwen3.8-Max, 3 August 2026
Smallest open model: Qwen3.8-27B (14 August 2026)
In Ollama: qwen3.8:27b, about 18 GB
Licence (27B): Apache 2.0

This page is the free part.

The course goes deeper on running and choosing models

₹999 in India$49 everywhere elseonce, for the whole course

The AI Engineering course covers running and choosing models across a run of lessons, not one page. These 4 alone are about 193 minutes of step-by-step reading, every one with code you run in the browser, all with a quiz.

Quantization: The Same Model in Fewer Bits
The same Qwen2.5 3B model stored three ways and measured on a laptop: 6.18 GB at 16 bits per weight, 3.29 GB at 8.5 and 1.93 GB at about 5. The 8-bit file kept the 16-bit file's top next word on 24 of 24 prompts and got the same 33 of 40 sums right; the 5-bit file kept 19 of 24, all 5 changes on near ties. Writing speed went from 11.1 to 18.3 to 28.4 tokens a second, while reading barely changed.
46 min · diagrams · code you run · quiz
Reading Is Fast, Writing Is Slow: Where a Model's Time Goes
Every Ollama reply reports how long the model spent loading, reading the prompt and writing the answer. Measured on a laptop with qwen2.5:3b over 35 calls in a shuffled order: the median reading speed was 194 tokens a second and the median writing speed 30. Even the slowest reading of any call was faster than the fastest writing of any call. A long reply costs far more time than a long prompt.
51 min · diagrams · code you run · quiz
Temperature and Sampling: How One Token Gets Picked
The model gives odds; a sampler picks one token. Measured with qwen2.5:3b on 3 prompts, 40 runs per setting: at temperature 0 every run wrote the same text; at 0.7 almost every run was different. For 'The capital of France is', 40 of 40 runs began with Paris at temperature 0 and only 5 of 40 at 1.5, where the text broke into other scripts and symbols. The same seed gave the same text every time.
50 min · diagrams · code you run · quiz
The Context Window: What Happens When a Prompt Does Not Fit
Measured with Ollama and qwen2.5:3b: a prompt of about 5,900 tokens sent with a 4,096-token window was cut to 2,050 tokens, the first 4 and the last 2,046, with no error in the reply. A code placed at the start was found 0 times in 5; at the end, 5 in 5. At three window sizes the kept part was exactly half the window plus 2. Without num_ctx, this laptop's window was 4,096.
46 min · diagrams · code you run · quiz

all part of

AI Engineering

204 lessons · about 211 hours · 10 free to read

₹999in India, by UPI

$49everywhere else, by PayPal

Get the AI Engineering course Read 10 lessons free first

You own this course

Continue with Quantization: The Same Model in Fewer Bits

pay once,
yours for life

What Qwen 3.8 is

“This also marks the first time we will open-source the weights of a Qwen-Max-class model”

Qwen team, Qwen3.8-Max blog post, 3 August 2026

Qwen is a family of large language models (LLMs) made by the Qwen team at Alibaba. An LLM is a program that has learned from a huge amount of text and can write, answer questions and write code. Qwen 3.8 is the newest version of the family.

Some Qwen models are open-weight. The weights are the numbers the model learned. When they are open, you can download them and, after that download, run the model on your own computer or server with no account and no internet connection. Other Qwen models are only offered as a service on Qwen's cloud.

Qwen 3.8 started on 3 August 2026 with Qwen3.8-Max. Max is Qwen's largest and strongest line, and its blog post says this is the first time the Qwen team opened the weights of a Max model. Over the next weeks, three open-weight models arrived on Hugging Face, the website where most open models are shared.

The Qwen 3.8 models, one by one

A model's size is counted in parameters, the learned numbers inside it. B means billion and T means trillion. Two of these models are mixture-of-experts (MoE) models. They are split into many small expert parts, and only a few experts work on each word, so only some parameters are active at a time. That makes them faster than their total size suggests, but you still need memory for all of them.

Model	Size	Open weights?	Licence
Qwen3.8-27B	27B, dense (all parameters used)	Yes, 14 Aug 2026	Apache 2.0
Qwen3.8-Flash-Next	125B, 6B active (MoE), plus a 51B lookup table	Yes, 26 Aug 2026	qwen-community-1.0
Qwen3.8-2.4T-A95B	2.4T, 95B active (MoE)	Yes, 12 Aug 2026	qwen3.8-max
Qwen3.8-Max	built on 2.4T-A95B	No, a cloud service	not applicable
Qwen3.8-Flash	built on Flash-Next	No, a cloud service	not applicable
Qwen3.8-Omni-Flash	not stated	No weights published	not applicable

From the official Hugging Face model cards and the QwenLM GitHub news lines, read 6 October 2026. The Hugging Face listing for author Qwen shows no Qwen 3.8 model smaller than 27B. Each open model also has an FP8 copy, which stores each number in 8 bits, half the size of the original 16.

Flash-Next is a special case: its card calls it an "experimental preview of the architecture that will underpin Qwen4". The hosted Qwen3.8-Flash, built on it, costs 0.15 US dollars per million input tokens and 0.47 per million output tokens, per the Qwen blog. A token is a small piece of text, often part of a word.

A bar chart on a log scale of parameters. Qwen3.8-27B: 27 billion, all active. Qwen3.8-Flash-Next: 125 billion total, 6 billion active. Qwen3.8-2.4T-A95B: 2,400 billion total, 95 billion active. Each bar shows total and active parameters and the licence. — Total size against the part that works on each word

What is new in Qwen 3.8

First, thinking. A thinking model writes out its reasoning first, then gives the answer. That helps on hard problems but costs time. The 27B card says thinking "is on by default and can be disabled per request". A setting called reasoning_effort picks how much it thinks: xhigh, medium or low. Qwen's own default is xhigh. Ollama 0.35.1 listed the levels as false, low, medium and xhigh, and uses medium unless you change it. Note that Ollama's help text mentions high, which this model does not list.

Second, images and video. The card describes the 27B as "a native vision-language model that understands images and videos". My test below gave it a bar chart to read.

Third, coding and agent work. Agents are AI programs that take steps on their own, such as running commands. For SWE-bench Pro and several other coding tests, Qwen used Claude Code, Anthropic's coding tool, as the harness. A harness is the program that drives the model through each task. Qwen also fixed some broken SWE-bench Pro tasks before running it. Terminal Bench used a different harness, Terminus.

The 27B uses the Qwen 3.5 design. It has 64 layers, which mix two kinds of attention, the part that lets each word look at the others. It can read 262,144 tokens as built, and about one million with extra settings.

A dot chart comparing Qwen3.8-27B with Qwen3.6-27B on six official benchmarks: SWE-bench Pro 61.7 against 53.5, Terminal Bench 2.1 73.0 against 63.4, LiveCodeBench v6 90.3 against 83.9, OSWorld-Verified 84.3 against 63.9, IFBench 79.5 against 69.1, GPQA Diamond 89.2 against 87.8. — Qwen3.8-27B against Qwen3.6-27B, from the official card

Can you run it on your computer?

To run a model, its weights must sit in memory: the graphics card's memory, or on a Mac, the memory that the processor and graphics chip share. The smallest Qwen 3.8 download in Ollama's library is qwen3.8:27b, listed at 18 GB; the copy I pulled on 6 October was 17.7 GB. Qwen itself publishes only 16-bit and 8-bit (FP8) files. Ollama's copy is a 4-bit version called Q4_K_M. It stores each number in about 4 to 5 bits instead of 16, so the file is far smaller.

My rented NVIDIA L4 graphics card has 23,034 MiB (about 24 GB) of memory. Ollama put the whole model on the card and used 18,004 MiB of it (about 18.9 GB) while answering. That was with a context window of 4,096 tokens, which is how much text the model can see at once. A bigger context window needs more memory; I did not measure how much more.

So a computer needs a little more than 18 GB of free memory for the model alone. My MacBook has 24 GB of memory shared by everything. My guess, which I did not test, is that this is too tight and part of the model would run on the slower processor. I also keep its disk free on purpose, so I rented a cloud machine. That is my judgement, not a number from Qwen: the card gives no memory figures, and its server examples spread the model over 4 graphics cards. Ollama also offers a version named 27b-mlx, built with MLX, Apple's own tool for running models on Apple Silicon Macs. I did not test it.

If 18 GB is too much, the newest small official Qwen models are from Qwen 3.5, the design Qwen 3.8 is built on. In Ollama, qwen3.5:4b-q4_K_M (Apache 2.0) is 3.3 GB and qwen3.5:2b-q4_K_M is 1.9 GB. They are smaller and less capable models, not Qwen 3.8. The bigger Flash-Next is in Ollama too, from 105 GB, which needs a machine with well over 100 GB of memory. Community-made 1-bit copies of the 27B exist, from 6.19 GB, but nobody official has measured their quality.

Memory bars. A 17.7 GB block for the qwen3.8:27b download. Below it, the memory our L4 card used while answering: 18,004 MiB (about 18.9 GB) of 23,034 MiB. Then three common memory sizes, 16 GB, 24 GB and 32 GB, with the model drawn inside each: it does not fit in 16, it is very tight in 24 shared on a Mac, it fits in 32. — 18 GB of model needs more than 18 GB of memory

Installing it and reading its details

I rented an AWS g6.2xlarge machine. It has 8 virtual CPUs, 32 GiB (about 34 GB) of memory and one NVIDIA L4 graphics card. I installed Ollama 0.35.1 with its official install script. Then ollama pull qwen3.8:27b downloaded 17.7 GB in about 131 seconds; that was inside Amazon's network, so a home connection will take much longer. The copy I got had the ID (digest) aaee06c39dcf. Ollama's web page listed a different ID, e118e4d12a70, for the same name (tag), and I do not know why.

ollama show prints what the model is. Read it from the top: 27.3 billion parameters, a context length of 262,144 tokens, the 4-bit Q4_K_M format, and that it needs Ollama 0.32.12 or newer. Under Capabilities it lists vision, tools (it can ask to call functions) and thinking, with the levels false, low, medium and xhigh. The sampling settings (temperature 1, top_p 0.95, top_k 20) match the card's advice for thinking mode.

$ ollama run --help
Run a model

Usage:
  ollama run MODEL [PROMPT] [flags]

Flags:
      --dimensions int          Truncate output embeddings to specified dimension (embedding models only)
      --format string           Response format (e.g. json)
  -h, --help                    help for run
      --hidethinking            Hide thinking output (if provided)
      --insecure                Use an insecure registry
      --keepalive string        Duration to keep a model loaded (e.g. 5m)
      --nowordwrap              Don't wrap words to the next line automatically
      --think string[="true"]   Enable thinking mode: true/false or high/medium/low for supported models
      --truncate                For embedding models: truncate inputs exceeding context length (default: true). Set --truncate=false to error instead
      --verbose                 Show timings for response

Environment Variables:
      OLLAMA_EDITOR                 Path to editor for interactive prompt editing (Ctrl+G)
      OLLAMA_HOST                   IP Address for the ollama server (default 127.0.0.1:11434)
      OLLAMA_NOHISTORY              Do not preserve readline history

$ ollama show qwen3.8:27b
  Model
    architecture        qwen35     
    parameters          27.3B      
    context length      262144     
    embedding length    5120       
    quantization        Q4_K_M     
    requires            0.32.12    

  Capabilities
    completion     
    vision         
    tools          
    thinking       
        levels     false, low, medium, xhigh    
        default    medium                       

  Projector
    architecture        clip       
    parameters          460.73M    
    embedding length    1152       
    dimensions          5120       

  Parameters
    draft_num_predict    4       
    min_p                0       
    presence_penalty     0       
    repeat_penalty       1       
    temperature          1       
    top_k                20      
    top_p                0.95    

  License
    Apache License               
    Version 2.0, January 2004    
    ...

Real output from the recording 03-show-qwen.cast, ANSI colour codes removed, otherwise unedited. The machine was a rented AWS g6.2xlarge; every recorded line went through a filter that removes addresses and account numbers.

A real terminal frame of ollama show qwen3.8:27b listing architecture qwen35, parameters 27.3B, context length 262144, quantization Q4_K_M, requires 0.32.12, and capabilities completion, vision, tools and thinking. — A frame from the recording, drawn by agg

A quick answer, with thinking off

First I asked a simple question. A flag is an extra option typed after a command: --think=false makes the model answer straight away, and --verbose makes Ollama print timings. One catch: --think=false only turns thinking off. As far as I can tell, Ollama keeps the model's stored settings (temperature 1, as ollama show listed) instead of the card's settings for no thinking.

The first time, the answer took 77.5 seconds in total: 43 seconds to load the model onto the card, 25 seconds to read the prompt, and 9 seconds to write the answer. I asked again with the model already loaded. That run is below: 4.0 seconds in total, and the answer came out at 17.77 tokens a second.

What we measured: On a rented NVIDIA L4 (about 24 GB), Ollama 0.35.1, qwen3.8:27b, 6 October 2026: cold first run (model not yet loaded) 77.54 s in total (43.03 s loading); warm run (model already in memory) 4.02 s in total, 65 tokens at 17.77 tokens/s. One run each, not a benchmark.

$ ollama run qwen3.8:27b --verbose --nowordwrap --think=false "In three short sentences, explain what a message queue is to a beginner."
⠙ ⠙ ⠹ A message queue acts like a digital waiting line where tasks or data are stored and passed along one at a time. It allows different parts of a system to communicate without needing to be ready or available at the exact same moment. This setup helps manage traffic, prevent crashes, and ensure that important information is processed in order.

total duration:       4.015594588s
load duration:        966.773µs
prompt eval count:    27 token(s)
prompt eval cached:   23 token(s)
prompt eval duration: 189.674ms
prompt eval rate:     21.09 tokens/s
eval count:           65 token(s)
eval duration:        3.656965s
eval rate:            17.77 tokens/s

$ ollama ps
NAME           ID              SIZE     PROCESSOR    CONTEXT    UNTIL              
qwen3.8:27b    aaee06c39dcf    17 GB    100% GPU     4096       4 minutes from now

Real output from the recording 04b-prompt-nothink-warm.cast, ANSI codes removed. The ⠙ marks are Ollama's spinner.

A thinking answer

Next, a small time problem with --think=low. Ollama shows the thinking between Thinking... and ...done thinking, then the answer. The model did the sum in two steps and got 12:15, which is right: 09:40 plus 2 hours is 11:40, plus 35 minutes is 12:15.

This run printed 95 tokens at 34.90 tokens a second, about twice as fast as the run above. One likely reason is a draft step that guesses a few tokens ahead. The card says the model was trained to predict several tokens at once. On our machine, the plain qwen3.8:27b tag listed draft_num_predict 4, and Ollama's server log reported speculative decoding. Simple, predictable text like sums lets more of those guesses be kept. Do not read these two runs as a speed test of thinking against no thinking.

$ ollama run qwen3.8:27b --verbose --nowordwrap --think=low "A train leaves at 09:40 and the trip takes 2 hours 35 minutes. When does it arrive? Answer in one line."
⠙ ⠹ ⠹ ⠸ Thinking...
The user wants to find the arrival time of a train.

Departure: 09:40
Duration: 2 hours 35 minutes

09:40 + 2 hours = 11:40
11:40 + 35 minutes = 12:15

So the arrival time is 12:15.
...done thinking.

The train arrives at 12:15.

total duration:       3.202382735s
load duration:        948.502µs
prompt eval count:    72 token(s)
prompt eval duration: 307.48ms
prompt eval rate:     234.16 tokens/s
eval count:           95 token(s)
eval duration:        2.722365s
eval rate:            34.90 tokens/s

Real output from the recording 05-prompt-think.cast, ANSI codes removed.

A real terminal frame of the thinking run: the model's thinking lines, 09:40 + 2 hours = 11:40, 11:40 + 35 minutes = 12:15, then done thinking and the answer The train arrives at 12:15, with the timing lines below. — The thinking run, from the recording

after “A thinking answer”

Want to try this yourself instead of only reading it? This lesson has code you run in your browser:

Thinking Step by Step: When It Helps, and What It Costs: The same questions asked two ways, answer only or think step by step and end with an Answer: line, on two small models. On 40 word problems, qwen2.5:3b went from 2 right to 37 and llama3.2:3b from 7 to 37. On two-digit sums, llama went from 23 to 37 while qwen moved from 30 to 32, a change within luck. On sorting messages, it lifted llama from 20 to 38 on a weak prompt and made no clear difference where the prompt was already good. Step-by-step replies had a median of 78 to 204 tokens, against 2 to 5 for direct ones.

57 min · diagrams · code you run · quiz

Get the AI Engineering course ₹999 in India · $49 elsewhere · pay onceor open the lesson

Continue in your course

Reading a chart from an image

To test vision, I drew a bar chart with known values using a short Python script: Mon 12, Tue 30, Wed 21, Thu 45, Fri 8. Then I passed the image to the model by putting its file path in the prompt.

It listed all five days with the right values. Reading the image took most of the time: 277 tokens of input in 4.1 seconds, then 29 tokens of answer in under a second. This was one easy chart and one try, so it shows that vision works, not how accurate it is.

What we measured: 5 of 5 values read correctly. Total 5.25 s; 277 prompt tokens (the image becomes tokens too); answer at 31.52 tokens/s.

$ cat chart.py
# chart.py: draw a simple bar chart with known values, for the image prompt test.
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt

days = ["Mon", "Tue", "Wed", "Thu", "Fri"]
orders = [12, 30, 21, 45, 8]
fig, ax = plt.subplots(figsize=(6, 4), dpi=100)
ax.bar(days, orders, color="#3b6ea5")
ax.set_title("Orders per day")
ax.set_ylabel("Orders")
fig.tight_layout()
fig.savefig("chart.png")
print("wrote chart.png", dict(zip(days, orders)))

$ python3 chart.py
wrote chart.png {'Mon': 12, 'Tue': 30, 'Wed': 21, 'Thu': 45, 'Fri': 8}

$ ollama run qwen3.8:27b --verbose --nowordwrap --think=false "Read this bar chart. List each day and its value, one per line. ./chart.png"
Added image './chart.png'
⠙ ⠹ ⠸ ⠼ ⠴ ⠦ ⠧ ⠇ ⠏ ⠋ ⠙ ⠹ ⠸ ⠼ ⠴ ⠦ ⠧ ⠇ ⠏ ⠋ ⠙ ⠹ ⠸ ⠼ ⠴ ⠦ ⠧ ⠇ ⠏ ⠋ ⠙ ⠹ ⠸ ⠼ ⠴ ⠦ ⠧ ⠇ ⠏ ⠋ ⠙ ⠹ ⠸ Mon: 12
Tue: 30
Wed: 21
Thu: 45
Fri: 8

total duration:       5.248751451s
load duration:        1.164188ms
prompt eval count:    277 token(s)
prompt eval duration: 4.145229s
prompt eval rate:     66.82 tokens/s
eval count:           29 token(s)
eval duration:        920.164ms
eval rate:            31.52 tokens/s

Real output from the recording 06-image.cast, ANSI codes removed. chart.py ran on the same machine.

The bar chart the model read: Orders per day. Mon 12, Tue 30, Wed 21, Thu 45, Fri 8. — The chart the model was given

Settings worth knowing

From the official 27B card and Ollama's docs, in plain words.

Thinking on (default). The card suggests temperature 1.0, top_p 0.95, top_k 20, min_p 0 and presence_penalty 0. Temperature controls how random the wording is; top_p and top_k limit which words it picks from.
Thinking off. The card suggests temperature 0.7, top_p 0.8, top_k 20, min_p 0 and presence_penalty 1.5. In Ollama, add --think=false (or send "think": false to its API, the way programs talk to Ollama), and set those values yourself, for example with /set parameter temperature 0.7 in a chat.
How hard it thinks. reasoning_effort can be xhigh (the card's default), medium or low. The card warns that lower effort does not always finish an agent's task sooner.
Repetition. presence_penalty pushes the model away from words it already used. If answers repeat, it can go from 0 to 2, but the card warns higher values may make the model mix languages.
Context window. Ollama used a context of 4,096 tokens on my card (shown by ollama ps). For long documents, raise the context in Ollama's settings, if you have the memory; a bigger context needs more of it.
Serving it for many users. The card recommends SGLang, vLLM or TokenSpeed, which are servers built for many users at once, rather than a desktop tool.

What I could and could not check

The lineup, licences, dates, settings and benchmark scores come from Qwen's own blog, GitHub and Hugging Face cards, read on 6 October 2026. The benchmarks are Qwen's numbers; I did not rerun them. The Ollama sizes come from ollama.com.

I ran only Qwen3.8-27B, the Ollama 4-bit copy. That was four runs of three prompts: the first prompt twice (cold, then warm) and the others once. They ran on one rented NVIDIA L4 GPU, because the cloud account could not start a large CPU-only machine that day. I did not run it on a Mac, so I have no Mac speeds. I did not test Flash-Next or the 2.4T model, tool calling, long documents or video. The machine cost about 27 US cents and was deleted afterwards.

Questions

What is Qwen 3.8?

The newest family of large language models from Alibaba's Qwen team, starting with Qwen3.8-Max on 3 August 2026. Three of its models have open weights: Qwen3.8-27B, Qwen3.8-Flash-Next and Qwen3.8-2.4T-A95B. Others, such as Qwen3.8-Max and Qwen3.8-Flash, are cloud services.

Is Qwen 3.8 open source?

Three models have open weights you can download. Only Qwen3.8-27B uses the Apache 2.0 licence. Flash-Next uses qwen-community-1.0 and the 2.4T model uses a qwen3.8-max licence, so read those before commercial use.

How do I run Qwen 3.8 with Ollama?

Install Ollama 0.32.12 or newer, the version that added the model. 0.32.15 fixed how Qwen 3.8 handles system messages, so a newer version is better; I used 0.35.1. Then run ollama run qwen3.8:27b. The download is about 18 GB and the model needs a little more than that in free graphics or shared memory. Add --think=false for quick answers or --think=low for short reasoning.

Is there a small Qwen 3.8 model?

No official one. The smallest official Qwen 3.8 model is 27B, 18 GB in Ollama. For small machines, the official Qwen 3.5 models are smaller, for example qwen3.5:4b-q4_K_M at 3.3 GB, but they are a different, less capable family.

How fast is Qwen3.8-27B?

In my test on one NVIDIA L4 GPU with Ollama, once loaded it wrote 17.8 to 34.9 tokens a second, depending on the text. Loading the model the first time took 43 seconds. These are single runs on one machine, not a benchmark, and a Mac will be different.

How do I turn thinking off in Qwen 3.8?

In Ollama, add --think=false to ollama run, or send "think": false to its API. With vLLM or SGLang, two servers for running models, the card says to send "chat_template_kwargs": {"enable_thinking": false}. On Qwen's cloud, send "enable_thinking": false.

Sources

What this explainer is based on, so you can check it.

Qwen3.8-Max: A New Bar for Coding and Cowork Qwen team, 3 August 2026
Qwen3.8-27B model card Qwen on Hugging Face
Qwen3.8-2.4T-A95B model card Qwen on Hugging Face
Qwen3.8-Flash-Next model card Qwen on Hugging Face
Qwen3.8-Flash-Next blog post Qwen team
Qwen3.8 on GitHub QwenLM on GitHub
qwen3.8 in the Ollama library (tags and sizes) Ollama
Ollama releases (0.32.12 adds Qwen 3.8 27B) Ollama on GitHub
Thinking in Ollama Ollama docs
Qwen3.5-4B model card Qwen on Hugging Face

Learn the parts behind it

Quantization: The Same Model in Fewer BitsThe same Qwen2.5 3B model stored three ways and measured on a laptop: 6.18 GB at 16 bits per weight, 3.29 GB at 8.5 and 1.93 GB at about 5. The 8-bit file kept the 16-bit file's top next word on 24 of 24 prompts and got the same 33 of 40 sums right; the 5-bit file kept 19 of 24, all 5 changes on near ties. Writing speed went from 11.1 to 18.3 to 28.4 tokens a second, while reading barely changed.46 min lesson Reading Is Fast, Writing Is Slow: Where a Model's Time GoesEvery Ollama reply reports how long the model spent loading, reading the prompt and writing the answer. Measured on a laptop with qwen2.5:3b over 35 calls in a shuffled order: the median reading speed was 194 tokens a second and the median writing speed 30. Even the slowest reading of any call was faster than the fastest writing of any call. A long reply costs far more time than a long prompt.51 min lesson Temperature and Sampling: How One Token Gets PickedThe model gives odds; a sampler picks one token. Measured with qwen2.5:3b on 3 prompts, 40 runs per setting: at temperature 0 every run wrote the same text; at 0.7 almost every run was different. For 'The capital of France is', 40 of 40 runs began with Paris at temperature 0 and only 5 of 40 at 1.5, where the text broke into other scripts and symbols. The same seed gave the same text every time.50 min lesson The Context Window: What Happens When a Prompt Does Not FitMeasured with Ollama and qwen2.5:3b: a prompt of about 5,900 tokens sent with a 4,096-token window was cut to 2,050 tokens, the first 4 and the last 2,046, with no error in the reply. A code placed at the start was found 0 times in 5; at the end, 5 in 5. At three window sizes the kept part was exactly half the window plus 2. Without num_ctx, this laptop's window was 4,096.46 min lesson Thinking Step by Step: When It Helps, and What It CostsThe same questions asked two ways, answer only or think step by step and end with an Answer: line, on two small models. On 40 word problems, qwen2.5:3b went from 2 right to 37 and llama3.2:3b from 7 to 37. On two-digit sums, llama went from 23 to 37 while qwen moved from 30 to 32, a change within luck. On sorting messages, it lifted llama from 20 to 38 on a weak prompt and made no clear difference where the prompt was already good. Step-by-step replies had a median of 78 to 204 tokens, against 2 to 5 for direct ones.57 min lesson Comparing Models Fairly: Same Test, Different ResultsFour small models measured on one laptop with the same tests: gemma3:4b got 33 of 40 sums right, qwen2.5:3b 30, llama3.2:3b 23, and qwen3:4b 0, because it reasoned in its reply and ran out of its 32-token budget. Given 1,024 tokens it finished 22 sums and got all 22 right, at a median 926 tokens a sum. Only one pairwise gap was clearly bigger than luck. qwen2.5 wrote fastest, 38.2 tokens a second; gemma3 wrote 20.2.47 min lesson

Go deeper than the news

This explainer sits on top of our AI Engineering course: 204 lessons on RAG, evals, agents, serving, security and MLOps, many built around a real experiment. 10 lessons are free to read, with no card needed.

course 2

AI Engineering

Take models from notebook to production, with labs on real models.

204 interactive lessons
Serving, pipelines, feature stores and monitoring
RAG, fine-tuning, evals and agents
Step-by-step diagrams and quizzes
Lifetime access and all future lessons
A separate purchase from System Design

Read 10 lessons free first

In India

₹999

Pay with UPI

Everywhere else

$49

Pay with PayPal

once · yours for life

ExplainedQwen 3.8Qwenopen-weight modelsOllamalocal LLMsthinking models

Qwen 3.8 explained: what is new, and running the 27B model with Ollama

Roni Das

14 min read

Published6 October 2026, 12:10 UTC

Two courses by the author of this page:

AI Engineering

204 lessons · 10 free to read

₹999 in India · $49 elsewhere, once

Get AI Engineering

You own this course

System Design

770 lessons · 18 free to read

₹499 in India · $49 elsewhere, once

Get System Design

You own this course

Every lesson in both courses has an AI tutor beside it. It reads the same lesson you are reading and answers your questions from it. Try the tutor

In short

Qwen 3.8 is a family, not one model. The open-weight models are Qwen3.8-27B, Qwen3.8-Flash-Next and Qwen3.8-2.4T-A95B; Qwen3.8-Max, Qwen3.8-Flash and Qwen3.8-Omni-Flash are services you call over the internet.
Only Qwen3.8-27B uses the Apache 2.0 licence, legal terms that allow use in paid products. The two big open models have their own Qwen licences, so read them before using them in a product.
The models think before they answer, unless you turn that off per request, and you can set how hard they think. The 27B model also reads images and videos.
The smallest Qwen 3.8 download in Ollama's library is qwen3.8:27b, about 18 GB. On a rented NVIDIA L4 GPU (graphics card) it wrote 18 to 35 tokens (small pieces of words) a second once loaded, depending on the text, in single runs. It read all 5 values from one simple bar chart.
On Qwen's own tests the 27B scores well above the older Qwen3.6-27B, for example 61.7 against 53.5 on SWE-bench Pro, a test of fixing real GitHub issues. Qwen ran these on its own setup, not on the 4-bit Ollama copy; I did not rerun them.

First model: Qwen3.8-Max, 3 August 2026
Smallest open model: Qwen3.8-27B (14 August 2026)
In Ollama: qwen3.8:27b, about 18 GB
Licence (27B): Apache 2.0

This page is the free part.

The course goes deeper on running and choosing models

₹999 in India$49 everywhere elseonce, for the whole course

Quantization: The Same Model in Fewer Bits
The same Qwen2.5 3B model stored three ways and measured on a laptop: 6.18 GB at 16 bits per weight, 3.29 GB at 8.5 and 1.93 GB at about 5. The 8-bit file kept the 16-bit file's top next word on 24 of 24 prompts and got the same 33 of 40 sums right; the 5-bit file kept 19 of 24, all 5 changes on near ties. Writing speed went from 11.1 to 18.3 to 28.4 tokens a second, while reading barely changed.
46 min · diagrams · code you run · quiz
Reading Is Fast, Writing Is Slow: Where a Model's Time Goes
Every Ollama reply reports how long the model spent loading, reading the prompt and writing the answer. Measured on a laptop with qwen2.5:3b over 35 calls in a shuffled order: the median reading speed was 194 tokens a second and the median writing speed 30. Even the slowest reading of any call was faster than the fastest writing of any call. A long reply costs far more time than a long prompt.
51 min · diagrams · code you run · quiz
Temperature and Sampling: How One Token Gets Picked
The model gives odds; a sampler picks one token. Measured with qwen2.5:3b on 3 prompts, 40 runs per setting: at temperature 0 every run wrote the same text; at 0.7 almost every run was different. For 'The capital of France is', 40 of 40 runs began with Paris at temperature 0 and only 5 of 40 at 1.5, where the text broke into other scripts and symbols. The same seed gave the same text every time.
50 min · diagrams · code you run · quiz
The Context Window: What Happens When a Prompt Does Not Fit
Measured with Ollama and qwen2.5:3b: a prompt of about 5,900 tokens sent with a 4,096-token window was cut to 2,050 tokens, the first 4 and the last 2,046, with no error in the reply. A code placed at the start was found 0 times in 5; at the end, 5 in 5. At three window sizes the kept part was exactly half the window plus 2. Without num_ctx, this laptop's window was 4,096.
46 min · diagrams · code you run · quiz

all part of

AI Engineering

204 lessons · about 211 hours · 10 free to read

₹999in India, by UPI

$49everywhere else, by PayPal

Get the AI Engineering course Read 10 lessons free first

You own this course

Continue with Quantization: The Same Model in Fewer Bits

pay once,
yours for life

What Qwen 3.8 is

“This also marks the first time we will open-source the weights of a Qwen-Max-class model”

Qwen team, Qwen3.8-Max blog post, 3 August 2026

The Qwen 3.8 models, one by one

Model	Size	Open weights?	Licence
Qwen3.8-27B	27B, dense (all parameters used)	Yes, 14 Aug 2026	Apache 2.0
Qwen3.8-Flash-Next	125B, 6B active (MoE), plus a 51B lookup table	Yes, 26 Aug 2026	qwen-community-1.0
Qwen3.8-2.4T-A95B	2.4T, 95B active (MoE)	Yes, 12 Aug 2026	qwen3.8-max
Qwen3.8-Max	built on 2.4T-A95B	No, a cloud service	not applicable
Qwen3.8-Flash	built on Flash-Next	No, a cloud service	not applicable
Qwen3.8-Omni-Flash	not stated	No weights published	not applicable

What is new in Qwen 3.8

Second, images and video. The card describes the 27B as "a native vision-language model that understands images and videos". My test below gave it a bar chart to read.

Can you run it on your computer?

Installing it and reading its details

$ ollama run --help
Run a model

Usage:
  ollama run MODEL [PROMPT] [flags]

Flags:
      --dimensions int          Truncate output embeddings to specified dimension (embedding models only)
      --format string           Response format (e.g. json)
  -h, --help                    help for run
      --hidethinking            Hide thinking output (if provided)
      --insecure                Use an insecure registry
      --keepalive string        Duration to keep a model loaded (e.g. 5m)
      --nowordwrap              Don't wrap words to the next line automatically
      --think string[="true"]   Enable thinking mode: true/false or high/medium/low for supported models
      --truncate                For embedding models: truncate inputs exceeding context length (default: true). Set --truncate=false to error instead
      --verbose                 Show timings for response

Environment Variables:
      OLLAMA_EDITOR                 Path to editor for interactive prompt editing (Ctrl+G)
      OLLAMA_HOST                   IP Address for the ollama server (default 127.0.0.1:11434)
      OLLAMA_NOHISTORY              Do not preserve readline history

$ ollama show qwen3.8:27b
  Model
    architecture        qwen35     
    parameters          27.3B      
    context length      262144     
    embedding length    5120       
    quantization        Q4_K_M     
    requires            0.32.12    

  Capabilities
    completion     
    vision         
    tools          
    thinking       
        levels     false, low, medium, xhigh    
        default    medium                       

  Projector
    architecture        clip       
    parameters          460.73M    
    embedding length    1152       
    dimensions          5120       

  Parameters
    draft_num_predict    4       
    min_p                0       
    presence_penalty     0       
    repeat_penalty       1       
    temperature          1       
    top_k                20      
    top_p                0.95    

  License
    Apache License               
    Version 2.0, January 2004    
    ...

A quick answer, with thinking off

$ ollama run qwen3.8:27b --verbose --nowordwrap --think=false "In three short sentences, explain what a message queue is to a beginner."
⠙ ⠙ ⠹ A message queue acts like a digital waiting line where tasks or data are stored and passed along one at a time. It allows different parts of a system to communicate without needing to be ready or available at the exact same moment. This setup helps manage traffic, prevent crashes, and ensure that important information is processed in order.

total duration:       4.015594588s
load duration:        966.773µs
prompt eval count:    27 token(s)
prompt eval cached:   23 token(s)
prompt eval duration: 189.674ms
prompt eval rate:     21.09 tokens/s
eval count:           65 token(s)
eval duration:        3.656965s
eval rate:            17.77 tokens/s

$ ollama ps
NAME           ID              SIZE     PROCESSOR    CONTEXT    UNTIL              
qwen3.8:27b    aaee06c39dcf    17 GB    100% GPU     4096       4 minutes from now

Real output from the recording 04b-prompt-nothink-warm.cast, ANSI codes removed. The ⠙ marks are Ollama's spinner.

A thinking answer

$ ollama run qwen3.8:27b --verbose --nowordwrap --think=low "A train leaves at 09:40 and the trip takes 2 hours 35 minutes. When does it arrive? Answer in one line."
⠙ ⠹ ⠹ ⠸ Thinking...
The user wants to find the arrival time of a train.

Departure: 09:40
Duration: 2 hours 35 minutes

09:40 + 2 hours = 11:40
11:40 + 35 minutes = 12:15

So the arrival time is 12:15.
...done thinking.

The train arrives at 12:15.

total duration:       3.202382735s
load duration:        948.502µs
prompt eval count:    72 token(s)
prompt eval duration: 307.48ms
prompt eval rate:     234.16 tokens/s
eval count:           95 token(s)
eval duration:        2.722365s
eval rate:            34.90 tokens/s

Real output from the recording 05-prompt-think.cast, ANSI codes removed.

after “A thinking answer”

Want to try this yourself instead of only reading it? This lesson has code you run in your browser:

57 min · diagrams · code you run · quiz

Get the AI Engineering course ₹999 in India · $49 elsewhere · pay onceor open the lesson

Continue in your course

Reading a chart from an image

To test vision, I drew a bar chart with known values using a short Python script: Mon 12, Tue 30, Wed 21, Thu 45, Fri 8. Then I passed the image to the model by putting its file path in the prompt.

What we measured: 5 of 5 values read correctly. Total 5.25 s; 277 prompt tokens (the image becomes tokens too); answer at 31.52 tokens/s.

$ cat chart.py
# chart.py: draw a simple bar chart with known values, for the image prompt test.
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt

days = ["Mon", "Tue", "Wed", "Thu", "Fri"]
orders = [12, 30, 21, 45, 8]
fig, ax = plt.subplots(figsize=(6, 4), dpi=100)
ax.bar(days, orders, color="#3b6ea5")
ax.set_title("Orders per day")
ax.set_ylabel("Orders")
fig.tight_layout()
fig.savefig("chart.png")
print("wrote chart.png", dict(zip(days, orders)))

$ python3 chart.py
wrote chart.png {'Mon': 12, 'Tue': 30, 'Wed': 21, 'Thu': 45, 'Fri': 8}

$ ollama run qwen3.8:27b --verbose --nowordwrap --think=false "Read this bar chart. List each day and its value, one per line. ./chart.png"
Added image './chart.png'
⠙ ⠹ ⠸ ⠼ ⠴ ⠦ ⠧ ⠇ ⠏ ⠋ ⠙ ⠹ ⠸ ⠼ ⠴ ⠦ ⠧ ⠇ ⠏ ⠋ ⠙ ⠹ ⠸ ⠼ ⠴ ⠦ ⠧ ⠇ ⠏ ⠋ ⠙ ⠹ ⠸ ⠼ ⠴ ⠦ ⠧ ⠇ ⠏ ⠋ ⠙ ⠹ ⠸ Mon: 12
Tue: 30
Wed: 21
Thu: 45
Fri: 8

total duration:       5.248751451s
load duration:        1.164188ms
prompt eval count:    277 token(s)
prompt eval duration: 4.145229s
prompt eval rate:     66.82 tokens/s
eval count:           29 token(s)
eval duration:        920.164ms
eval rate:            31.52 tokens/s

Real output from the recording 06-image.cast, ANSI codes removed. chart.py ran on the same machine.

Settings worth knowing

From the official 27B card and Ollama's docs, in plain words.

Thinking on (default). The card suggests temperature 1.0, top_p 0.95, top_k 20, min_p 0 and presence_penalty 0. Temperature controls how random the wording is; top_p and top_k limit which words it picks from.
Thinking off. The card suggests temperature 0.7, top_p 0.8, top_k 20, min_p 0 and presence_penalty 1.5. In Ollama, add --think=false (or send "think": false to its API, the way programs talk to Ollama), and set those values yourself, for example with /set parameter temperature 0.7 in a chat.
How hard it thinks. reasoning_effort can be xhigh (the card's default), medium or low. The card warns that lower effort does not always finish an agent's task sooner.
Repetition. presence_penalty pushes the model away from words it already used. If answers repeat, it can go from 0 to 2, but the card warns higher values may make the model mix languages.
Context window. Ollama used a context of 4,096 tokens on my card (shown by ollama ps). For long documents, raise the context in Ollama's settings, if you have the memory; a bigger context needs more of it.
Serving it for many users. The card recommends SGLang, vLLM or TokenSpeed, which are servers built for many users at once, rather than a desktop tool.

What I could and could not check

Questions

What is Qwen 3.8?

Is Qwen 3.8 open source?

How do I run Qwen 3.8 with Ollama?

Is there a small Qwen 3.8 model?

How fast is Qwen3.8-27B?

How do I turn thinking off in Qwen 3.8?

Sources

What this explainer is based on, so you can check it.

Qwen3.8-Max: A New Bar for Coding and Cowork Qwen team, 3 August 2026
Qwen3.8-27B model card Qwen on Hugging Face
Qwen3.8-2.4T-A95B model card Qwen on Hugging Face
Qwen3.8-Flash-Next model card Qwen on Hugging Face
Qwen3.8-Flash-Next blog post Qwen team
Qwen3.8 on GitHub QwenLM on GitHub
qwen3.8 in the Ollama library (tags and sizes) Ollama
Ollama releases (0.32.12 adds Qwen 3.8 27B) Ollama on GitHub
Thinking in Ollama Ollama docs
Qwen3.5-4B model card Qwen on Hugging Face

Learn the parts behind it

Go deeper than the news

course 2

AI Engineering

Take models from notebook to production, with labs on real models.

204 interactive lessons
Serving, pipelines, feature stores and monitoring
RAG, fine-tuning, evals and agents
Step-by-step diagrams and quizzes
Lifetime access and all future lessons
A separate purchase from System Design

Read 10 lessons free first

In India

₹999

Pay with UPI

Everywhere else

$49

Pay with PayPal

once · yours for life