Back to Gap Map Model components / Base / pretrained models

Qwen Base

open weights / Overall score: 4.0(strong)

Alibaba's pretrained Qwen base models, published for fine-tuning and research rather than chat. The current base generation is Qwen3.5, released in February 2026 in five sizes from 0.8B to a 35B mixture of experts with 3B active parameters. Qwen has published no base checkpoints for Qwen 3.6 or 3.8, whose open-weight models ship post-trained only.

Covers Qwen's base checkpoints only. The post-trained models are in qwen-instruct, the same split as llama and llama-instruct; the aliases for the Qwen 3.6 to 3.8 releases point here from before that split.

Openness

3 high confidence

3.0

weights: open(ungated on the Hub: the five Qwen3.5 Base checkpoints, 0.8B to 35B-A3B)
data: closed(no corpus or data card published
code: partial(usage, deployment and fine-tuning pointers in QwenLM/Qwen3.8, which covers the Qwen3.5 series
license: Apache-2.0(OSI

Qwen3.5's base models download from Hugging Face under Apache-2.0 without a gate, in five sizes. The pretraining corpus is not released and only usage code is published, so this is open weights rather than open source.

https://huggingface.co/api/models/Qwen/Qwen3.5-0.8B-Base recorded 2026-10-09
Hub record for Qwen/Qwen3.5-0.8B-Base: "private": false, "gated": false, card license "apache-2.0", created 2026-02-28.
https://huggingface.co/api/models/Qwen/Qwen3.5-2B-Base recorded 2026-10-09
Hub record for Qwen/Qwen3.5-2B-Base: "private": false, "gated": false, card license "apache-2.0", created 2026-02-28.
https://huggingface.co/api/models/Qwen/Qwen3.5-35B-A3B-Base recorded 2026-10-09
Hub record for the largest base checkpoint: "private": false, "gated": false, card license "apache-2.0", created 2026-02-24; the card front matter names no datasets.
https://huggingface.co/api/models/Qwen/Qwen3.5-4B-Base recorded 2026-10-09
Hub record for Qwen/Qwen3.5-4B-Base: "private": false, "gated": false, card license "apache-2.0", created 2026-02-27.
https://huggingface.co/api/models/Qwen/Qwen3.5-9B-Base recorded 2026-10-09
Hub record for Qwen/Qwen3.5-9B-Base: "private": false, "gated": false, card license "apache-2.0", created 2026-02-26.
https://huggingface.co/Qwen/Qwen3.5-35B-A3B-Base/raw/main/LICENSE recorded 2026-10-09
The Apache License, "Version 2.0, January 2004", in full.
https://huggingface.co/Qwen/Qwen3.5-35B-A3B-Base/raw/main/README.md recorded 2026-10-09
The card says the repository "contains model weights and configuration files for the pre-trained only model" and lists "Training Stage: Pre-training"; its intended uses are "fine-tuning, in-context learning experiments, and other research or development purposes, not direct interaction". It publishes no corpus.
https://raw.githubusercontent.com/QwenLM/Qwen3.8/main/README.md recorded 2026-10-09
The repository "of the Qwen3.5 open model series, including Qwen3.5, Qwen3.6, and the latest Qwen3.8". Its sections cover quickstart, local use and deployment, and for fine-tuning it points to third-party frameworks (Unsloth, Swift, Llama-Factory); no training pipeline is published.

Adoption

4 high confidence

4.0

Adoption is measured as Hugging Face downloads summed across the seven declared base checkpoints, the five Qwen3.5 Base sizes and the Qwen2.5 72B and 7B bases. The Qwen2.5-7B base still draws the most downloads of the seven.

https://huggingface.co/api/models/Qwen/Qwen2.5-72B recorded 2026-10-09
"downloads":16295 over the trailing 30 days for Qwen/Qwen2.5-72B.
https://huggingface.co/api/models/Qwen/Qwen2.5-7B recorded 2026-10-09
"downloads":781894 over the trailing 30 days for Qwen/Qwen2.5-7B.
https://huggingface.co/api/models/Qwen/Qwen3.5-0.8B-Base recorded 2026-10-09
"downloads":479031 over the trailing 30 days for Qwen/Qwen3.5-0.8B-Base.
https://huggingface.co/api/models/Qwen/Qwen3.5-2B-Base recorded 2026-10-09
"downloads":304328 over the trailing 30 days for Qwen/Qwen3.5-2B-Base.
https://huggingface.co/api/models/Qwen/Qwen3.5-35B-A3B-Base recorded 2026-10-09
"downloads":99083 over the trailing 30 days for Qwen/Qwen3.5-35B-A3B-Base.
https://huggingface.co/api/models/Qwen/Qwen3.5-4B-Base recorded 2026-10-09
"downloads":648400 over the trailing 30 days for Qwen/Qwen3.5-4B-Base.
https://huggingface.co/api/models/Qwen/Qwen3.5-9B-Base recorded 2026-10-09
"downloads":305121 over the trailing 30 days for Qwen/Qwen3.5-9B-Base.

Capability

4 medium confidence

4.0

Artificial Analysis does not evaluate base checkpoints, so this record is read on its post-trained sibling, as llama is read on Llama 4 Maverick. Qwen3.5 35B A3B, the chat model built on the largest Qwen3.5 Base checkpoint, scores 19 on Artificial Analysis's Intelligence Index, well above the median of 8 for open-weight models of its size, and between the base records read at 10 and at 39.

https://artificialanalysis.ai/models/llama-4-maverick recorded 2026-10-10
The llama record's reading on the same index version, fetched the same day: "Llama 4 Maverick scores 10 on the Artificial Analysis Intelligence Index", unchanged since 2026-09-29, so the two readings are on one scale.
https://artificialanalysis.ai/models/qwen3-5-35b-a3b recorded 2026-10-10
Artificial Analysis page on Intelligence Index v4.3.2: "Qwen3.5 35B A3B (Reasoning) scores 19 (estimated) on the Artificial Analysis Intelligence Index, placing it well above average among other open weight models of similar size (median: 8)".
https://artificialanalysis.ai/models/qwen3-5-397b-a17b recorded 2026-10-10
Artificial Analysis page for the larger Qwen3.5 post-trained model, which has no released base checkpoint: "Qwen3.5 397B A17B (Reasoning) scores 18 on the Artificial Analysis Intelligence Index", so the 35B A3B reading is the generation's strongest.

Verified 2026-10-09, except some scores, still awaiting re-confirmation