AI Potluck
Contribute
Back to Gap Map Model components / Fine-tuned / chat models

Qwen 3.8

Alibaba Cloud
open weights / Overall score: 4.3(strong)

Alibaba's post-trained Qwen chat models. The current release, Qwen 3.8 from August 2026, ships three open-weight models that think by default and can answer directly on request: a dense 27B under Apache-2.0, Qwen3.8-Flash-Next, a 125B mixture of experts with 6B active parameters, and the 2.4-trillion-parameter Qwen3.8-2.4T-A95B, which is also the base of the hosted Qwen3.8-Max. From Qwen 3.5 on, Alibaba publishes its chat checkpoints under the plain release name rather than as Instruct models.

Covers Qwen's post-trained chat models; the base checkpoints are in qwen, the same split as llama-instruct and llama. Qwen2.5-14B-Instruct, the release this record was first written for, stays declared.

Openness

3 high confidence
3.0
weights
open(ungated on the Hub: Qwen3.8-27B, Qwen3.8-Flash-Next and the Qwen3.8-2.4T-A95B flagship)
data
closed(no post-training data published
code
partial(usage and deployment instructions in QwenLM/Qwen3.8 and the model cards
license
Qwen-Community-License-1.0(Qwen3.8-Flash-Next

Qwen 3.8 ships its dense 27B chat model under Apache-2.0 and its Flash-Next and flagship models under Qwen's own licenses, which require model-as-a-service and AI coding or office assistant businesses to get a separate license before commercial use. Both licenses also require a product with more than 100 million monthly active users or $20 million in monthly revenue to display the model's name. The post-training data is not released and only usage code is published.

  • https://huggingface.co/api/models/Qwen/Qwen3.8-2.4T-A95B recorded 2026-10-09

    Hub record for the flagship: "private": false, "gated": false, card license "other" with license_name "qwen3.8-max", created 2026-08-08; the card front matter names no datasets.

  • https://huggingface.co/api/models/Qwen/Qwen3.8-27B recorded 2026-10-09

    Hub record for Qwen3.8-27B: "private": false, "gated": false, card license "apache-2.0", created 2026-08-05; the card front matter names no datasets.

  • https://huggingface.co/api/models/Qwen/Qwen3.8-Flash-Next recorded 2026-10-09

    Hub record for Qwen3.8-Flash-Next: "private": false, "gated": false, card license "other" with license_name "qwen-community-1.0", created 2026-08-24; the card front matter names no datasets.

  • https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/LICENSE recorded 2026-10-11

    Titled "Qwen3.8-Max License". A licensee that "conducts a Model as a Service or AI Work Assistant business" with aggregate revenue over "US$50,000,000" in twelve months must "obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose", internal use excepted. Clause 1 requires a commercial product with "more than 100,000,000 monthly active users or US$ 20,000,000 (or equivalent in other currencies) monthly revenue" to display the model name.

  • https://huggingface.co/Qwen/Qwen3.8-27B/raw/main/LICENSE recorded 2026-10-09

    The Apache License, "Version 2.0, January 2004", in full.

  • https://huggingface.co/Qwen/Qwen3.8-Flash-Next/raw/main/LICENSE recorded 2026-10-11

    Titled "Qwen Community License 1.0". Any licensee that "conducts a Model as a Service or AI Work Assistant business" must "obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose", with no revenue floor; internal use is excepted. Clause 1 requires a commercial product with "more than 100,000,000 monthly active users or US$ 20,000,000 (or equivalent in other currencies) monthly revenue" to display the model name.

  • https://raw.githubusercontent.com/QwenLM/Qwen3.8-Flash-Next/main/tech_report.pdf recorded 2026-10-09

    The Qwen3.8-Flash-Next technical report. Its sections are Model Architecture, Optimization and Evaluation; none publishes the post-training data or its mixture.

  • https://raw.githubusercontent.com/QwenLM/Qwen3.8/main/README.md recorded 2026-10-09

    The Qwen3.8 repository README covers introduction, benchmarks, quickstart, local use and deployment; for fine-tuning it points to third-party frameworks (Unsloth, Swift, Llama-Factory), and for the license it points to "the license file released with the model weights". No post-training pipeline or data is published.

Adoption

5 high confidence
5.0

Adoption is measured as Hugging Face downloads summed across the eight declared chat checkpoints, from Qwen2.5-14B-Instruct through the three Qwen 3.8 models. Qwen3.8-27B alone accounts for about two fifths of the total.

Capability

4 high confidence
4.0

This score reads Qwen 3.8, the current release of Qwen's chat models. On the Artificial Analysis index its flagship sits a point behind Gemini 3.8 Flash and well behind the closed leaders, and even the dense 27B model places above Gemini 3.1 Pro, so it scores level with Gemini Flash.

  • https://artificialanalysis.ai/leaderboards/models recorded 2026-10-09

    The leaderboard's embedded model records, ranked by intelligenceIndex, put Claude Opus 5.5 (Max) first at 57.62 and Grok 4.7 (Xhigh) at 46.45 among models not marked deprecated; Gemini 3.8 Flash (High) follows at 40.93, Qwen3.8 2.4T A95B at 39.89, Qwen3.8-Flash-Next at 39.82, Qwen3.8 27B (Xhigh) at 33.70, K2 Horizon 375B A23B at 30.5 and Gemini 3.1 Pro Preview at 29.72. The digest drifts between fetches because the page carries live pricing and throughput figures.

  • https://artificialanalysis.ai/models/qwen3-8-2-4t-a95b recorded 2026-10-09

    "Qwen3.8 2.4T A95B scores 40 on the Artificial Analysis Intelligence Index, placing it well above average among other open weight models of similar size (median: 18)".

  • https://artificialanalysis.ai/models/qwen3-8-27b recorded 2026-10-09

    "Qwen3.8 27B (Xhigh) scores 34 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/qwen3-8-flash-next recorded 2026-10-09

    "Qwen3.8-Flash-Next scores 40 on the Artificial Analysis Intelligence Index, placing it well above average among other open weight models of similar size (median: 18)".

  • https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/README.md recorded 2026-10-09

    The card says the repository "contains model weights and configuration files for the post-trained model", lists "Training Stage: Pre-training & Post-training", and introduces "Qwen3.8, the most capable generation in the Qwen open-model family to date".

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