Catalogue

The Qwen family:
branches and generations

3.8 open Coder 3.7 — 3

Inside the family the lines diverge by purpose, not only by size. Below is what each branch contains, how it differs from its neighbour, and what to look for in a model card before downloading.

Compact models→ Back to overview↗
Branches

What each line contains

Four models released in August—September 2026: two closed flagships and two open ones. Context is a million tokens for all of them, including the cheap flash.

Model Parameters Input Input, ¥/1M Output, ¥/1M
qwen3.8-max2.4T MoEtext, image, video1236
qwen3.8-flashmultimodal, fasttext, image, video0.82.7
qwen3.8-2.4t-a95b2.4T MoE / 95B activetext only1236
qwen3.8-27b27B densetext, image, video312

Prices are for the Beijing region, per million tokens; Singapore is roughly a quarter more expensive. A cache hit is an order of magnitude cheaper than regular input in every case: 0.1 ¥ on flash.

Structured output and tool calling are supported by all four models. Batch inference is max-only and Beijing-only; fine-tuning is not supported anywhere.

The open 3.8 line is the same flagship MoE released as weights, plus a compact dense model with vision. Both are available in two regions.

Model Architecture Input Context Reasoning chain
qwen3.8-2.4t-a95bsparse MoE, hybrid attentiontext only1M131,072
qwen3.8-27bdense, native visiontext, image, video1M262,144

Published results for the open flagship, from the model card:

Benchmark What it measures Score
GPQA DiamondGraduate-level science questions92.6
PaperBenchReproducing research papers93.0
OSWorldWorking inside a real operating system86.1
BabyVisionVisual understanding82.0
CodeArenaCode competition4th place

The scores come from the release documentation, not from our own runs. The weight footprint rules out a home deployment: this is a server story.

The code and maths branch: separate models with different post-training. For local agents a thirty-billion MoE is often enough, while editor autocomplete is served even by a 1.5B model from earlier generations.

Model Profile Access
qwen3-coder-nextAgentic scenarios, long contextservice
qwen3-coder-plusGeneration, editing, reviewservice
qwen3-coder-flashBulk operations, high throughputservice
qwen3-coder-480b-a35bThe open flagship for codeweights
qwen3-coder-30b-a3bOpen MoE: a speed/quality compromiseweights
qwen-math-plus / turboStep-by-step solutions, checking derivationsservice

Coder models almost always beat base models of the same size on structured output — that shows up clearly in data extraction.

Generations 3.7, 3.6, 3.5 and 3 keep working and remain the most practical choice for local deployment: quantised builds and tooling already exist for them.

Model Size Why it is useful
qwen3.7-maxclosedFlagship of the previous generation
qwen3.6-27b27B denseThe previous compact model with vision
qwen3.6-35b-a3b35B / 3BA fast mid-range MoE
qwen3.5-397b-a17b397B / 17BA large MoE with moderate activation
qwen3.5-122b-a10b122B / 10BA server MoE built for load
qwen3.5-35b-a3b35B / 3BThe workhorse for a single card
qwen3-next-80b-a3b80B / 3BExperimental line, cheap activation
qwen3-235b-a22b235B / 22BThe open flagship of generation 3
qwen3-30b-a3b30B / 3BThe popular compromise for a local server
qwen3-8b / 14b / 32b8—32BDense models: predictable speed
qwen3-0.6b … 4b0.6—4BExtraction and routing on CPU
qwq-plus—Reasoning in a hosted variant

Older generations do not "expire" automatically: they have more ready-made quantised builds and a lower cost to run.

Narrow models for one specific operation — you choose them for predictability and price on a bulk task, not for cleverness.

Model Task Note
qwen-mt-plus / flash / liteTranslationThree price tiers for different volumes
qwen-doc-turboDocument parsingCheaper than a general model on its home turf
qwen-longVery long textsThe historic line for long context
qwen-deep-researchSearch and gathering materialA service layered over search
qwen-plus-characterRole-play scenariosSeparate post-training for dialogue
tongyi-intent-detect-v3Intent detectionA tiny model for routing

For data extraction, narrow models are often better value than general ones: they are steadier on their own task class and cheaper at volume.

Dense and MoE: what changes in practice

A dense model computes all of its weights for every token. An MoE keeps many "experts" in memory but activates only a few of them. Hence the main consequence: you save on computation, not on memory. The open flagship activates roughly 95 billion parameters, yet 2.4 trillion have to be resident in memory.

So an MoE wins where the bottleneck is inference speed, and loses where memory is short. For a home node this means one simple thing: a dense model up to 8–14B behaves more predictably than an MoE that keeps spilling to disk.

Do not look at the number in the name; look at how many weights must be resident in memory at the same time.
Glossary

Terms from model cards

Hybrid attention

A mix of full and sparse attention: this keeps a million tokens of context computationally affordable.

Active parameters

How many weights take part in computing one token in an MoE. Marked by the second part of the name: A95B, A3B.

Reasoning chain

The limit on "thinking" before answering. Up to 262,144 tokens on 3.8, half that on the open flagship.

Context cache

Reusing an already computed prefix. A hit costs several times less than regular input.

Quantisation

Packing weights: Q4_K_M is the compromise, Q5/Q6 favour quality, AWQ and GPTQ are for server inference.

Structured output

Generation constrained by a schema or grammar: the model physically cannot emit an invalid answer.

Primary sources

Where to read the model cards

Weights and model cards — Hugging Face Code, reports and recipes — GitHub Our write-up on the compact models

QwenNotes is an independent site. We do not publish weights, we do not hand out access and we are not affiliated with the rights holder. Check specs and prices against the documentation: they change more often than our notes are published.