Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › Qwen family

Qwen-VL

Hugging Face ↗ Follow changes

Qwen Not declared out 3 years ago

Text in, text out.

Verified 2026-10-25 Rank #562 in LLMs · score 74.7 method v1 Sources 3 Downloads 30d 8.2k · #520 of 619 tracked

Can I use it?Not declaredNo license declared on the model cardlicence file
Do people use it?8.2kdownloads, 30 days#520 of 619 we track · below the median+4 likes, downloads +10% this week294 likesHugging Face

Worth reading

Picked by rule from 35 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Latest · 21 days ago

Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s

Strata runs the 125-billion-parameter Qwen3.8-Flash-Next model on consumer hardware under an MIT license with no cloud dependencies.

  • RX 6000 cards lack a dedicated installer and must fetch the engine automatically after installing hip-rdna3 or amd-rx7900.
  • Strata never splits a single model across NVIDIA and AMD graphics cards, even if multiple cards of the same make are present.
github.com 2026-10-04 924 HN pts · 422 comments 3.4k words about the Qwen family family

25 best of 35 tracked

The 2 picks above, plus these 23. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleQwen Image 2.1qwen.ai · 2026-09-20 · 740 HN pts · official
  2. ArticleQwen3.8-Flash-Nextqwen.ai · 2026-08-26 · 704 HN pts · official
  3. ArticleQwen3.8-Max: A New Bar for Coding and Coworkqwen.ai · 2026-08-03 · 1,124 HN pts · official
  4. ArticleQwen 3.8 27B is excellent, but it defaults to overthinking thingssimonwillison.net · 2026-08-16 · 802 HN pts
  5. ArticleQwen 3.8 27B available on Cerebras at 1500 tokens/sinference-docs.cerebras.ai · 2026-09-03 · 691 HN pts
  6. ArticleQwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledgeqwen.ai · 2026-07-21 · 571 HN pts · official
  7. DiscussionQwen 3.8twitter.com · 2026-07-19 · 961 HN pts
  8. ArticleQwen3.7-Max: The Agent Frontierqwen.ai · 2026-05-20 · 721 HN pts · official
  9. CodeShow HN: Qwen Scribe – local transcription and dictation for Apple Silicongithub.com · 2026-07-29
  10. ArticleLocal Qwen isn't a worse Opus, it's a different toolblog.alexellis.io · 2026-06-18
  11. ArticleQwen3.6-27B: Flagship-Level Coding in a 27B Dense Modelqwen.ai · 2026-04-22 · official
  12. ArticleQwen3.6-35B-A3B: Agentic coding power, now open to allqwen.ai · 2026-04-16 · official
  13. ArticleQwen3.6-Max-Preview: Smarter, Sharper, Still Evolvingqwen.ai · 2026-04-20 · official
  14. ArticleQwen3.6-Plus: Towards real world agentsqwen.ai · 2026-04-02 · official
  15. ArticleSomething is afoot in the land of Qwensimonwillison.net · 2026-03-04
  16. ArticleHow to run Qwen 3.5 locallyunsloth.ai · 2026-03-07
  17. ArticleQwen3-Coder-Nextqwen.ai · 2026-02-03 · official
  18. ArticleQwen3-Max-Thinkingqwen.ai · 2026-01-26 · official
  19. ArticleQwen3-TTS family is now open sourced: Voice design, clone, and generationqwen.ai · 2026-01-22 · official
  20. ArticleQwen3-Nextqwen.ai · 2025-09-12 · official
  21. ArticleQwen-Image: Crafting with native text renderingqwenlm.github.io · 2025-08-04 · official
  22. ArticleQwen3-VLqwen.ai · 2025-09-23 · official
  23. ArticleQwen VLo: From “Understanding” the World to “Depicting” Itqwenlm.github.io · 2025-06-27 · official

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves.

  • As a service0no API provider listed
  • On own hardware0quantized builds
  • Built on43 fine-tunes · 1 adapters · 0 merges
  • In apps0Hugging Face Spaces

The story so far

Weights, every dated article, and repo commits on one rail.

10 dated articles and 0 repo commits so far

Teal is the weights (2023-08-18), amber triangles are 10 dated articles (each links to it).

SepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctQwen Image 2.1 · 2026-09-20Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s · 2026-10-04Qwen3.8-Flash-Next · 2026-08-26Qwen3.8-Max: A New Bar for Coding and Cowork · 2026-08-03Qwen 3.8 27B is excellent, but it defaults to overthinking things · 2026-08-16Qwen 3.8 27B available on Cerebras at 1500 tokens/s · 2026-09-03Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledge · 2026-07-21Qwen 3.8 · 2026-07-19Qwen 3.6 27B is the sweet spot for local development · 2026-06-29Qwen3.7-Max: The Agent Frontier · 2026-05-20Weights 2023-08-18 · 2023-08-18Weights 2023-08-18now
SepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctQwen Image 2.1 · 2026-09-20Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s · 2026-10-04Qwen3.8-Flash-Next · 2026-08-26Qwen3.8-Max: A New Bar for Coding and Cowork · 2026-08-03Qwen 3.8 27B is excellent, but it defaults to overthinking things · 2026-08-16Qwen 3.8 27B available on Cerebras at 1500 tokens/s · 2026-09-03Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledge · 2026-07-21Qwen 3.8 · 2026-07-19Qwen 3.6 27B is the sweet spot for local development · 2026-06-29Qwen3.7-Max: The Agent Frontier · 2026-05-20Weights 2023-08-18 · 2023-08-18Weights 2023-08-18now

What people built on it

4 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

4 derivatives on the Hub, most of them fine-tunes

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

3 fine-tunes (Hugging Face model tree)1 adapter (Hugging Face model tree)3 fine-tunes1 adapter
3 fine-tunes (Hugging Face model tree)1 adapter (Hugging Face model tree)3 fine-tunes1 adapter

All variants →

Reference

Everything else, folded: the full tables, every source.

Use-case evidence

1 use case

What it is good at 1 use case

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

12 fields

Listed in LLMs

Facts

Pipeline tagtext-generation
Librarytransformers
ArchitecturesQWenLMHeadModel
Model typeqwen
MoEno
PapersarXiv:2308.12966
LicenseNot declared
Hub snapshot2026-10-25

All variants

4 variants

Lineage

Original model Family: Qwen/Qwen 4 derivatives on the Hub

finetune: 3adapter: 1

Top variants (4 shown, ranked by downloads)

Relation
4 rows
Variant Relation Quant Params Band License Author Likes Downloads
Hatim2221/Mubsir-Qwen-2B-VL finetune — 2.21B M apache-2.0 Hatim2221 5 3.2k
Sayeem26s/Chartqwen finetune — — — apache-2.0 Sayeem26s 1 0
leonardtzf/anim3bg-visual-novel-background-v2 adapter — — — apache-2.0 leonardtzf 0 0
fluorites/SemiEarth finetune — — — apache-2.0 fluorites 0 0

Articles and coverage

35 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official5

Qwen Image 2.1
qwen.ai2026-09-20officialauthorityHN front pageabout the Qwen familyHN 740 pts · 199 comments1,216 words
Qwen-Image-2.1 is an open-source image model released on September 20, 2026, that unifies text-to-image generation and image editing. The model uses a visual generation component with 32 Single-Stream DiT layers and 7B parameters, employing a mixed-granularity attention architecture to optimize inference efficiency. It supports native transparency and allows users to use up to 10 reference images for generating coherent compositions.written, reviewed, cited
  • The model uses KV cache reuse to compute and cache input images and editing instructions in the first step.
  • Qwen-Image-2.1 accepts the original image and a separate mask as two inputs to preserve complete input during local editing.
  • Text rendering considers type styles, layout, and their relationship to the overall composition, not just content.
Qwen3.8-Flash-Next
qwen.ai2026-08-26officialauthorityHN front pageabout the Qwen familyHN 704 pts · 233 comments3,581 words
Qwen3.8-Flash-Next is a 125B-parameter multimodal MoE model with 51B N-gram embeddings and 6B activated parameters per token. It natively supports 262,144 tokens of context and is extensible to 1,000,000 tokens with YaRN. The production version on QwenCloud costs 0.15 USD per million input tokens and 0.47 USD per million output tokens.written, reviewed, cited
  • The model uses a Muon optimizer refined around orthogonalization accuracy and the division of labour between Muon and AdamW.
  • Qwen Sparse Attention compresses the sequence into micro-blocks to estimate context importance at the block level before selecting regions.
  • Gated Residual expands the original single residual stream into four parallel branches to dynamically control reads and writes.
Qwen3.8-Max: A New Bar for Coding and Cowork
qwen.ai2026-08-03officialauthorityHN front pageabout the Qwen familyHN 1,124 pts · 612 comments11,334 words
Qwen 3.8-Max is a 2.4 trillion parameter model with 95B active parameters available via QwenCloud, featuring open weights releasing next week. The model self-evolves through feedback loops to complete complex tasks end-to-end with greater reliability. Over 16 days of fully autonomous AI operation, the repository accumulated 265 commits, 127 PRs, and 151 issues.written, reviewed, cited
  • The autonomous coding harness routes abnormal states back to relevant issues for fixes and re-verification.
  • The model beats the paper's own approach with a +2.7-point gain on the AIME24 math benchmark.
Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledge
qwen.ai2026-07-21officialauthorityHN front pageabout the Qwen familyHN 571 pts · 217 comments1,381 words
Qwen3.7-Max: The Agent Frontier
qwen.ai2026-05-20officialauthorityHN front pageabout the Qwen familyHN 721 pts · 290 comments3,821 words

Analysis4

Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s
github.com2026-10-04authorityHN front pageabout the Qwen familyHN 924 pts · 422 comments3,398 words
Strata runs the 125-billion-parameter Qwen3.8-Flash-Next model on consumer hardware under an MIT license with no cloud dependencies. Users need at least 32 GB of combined RAM and graphics memory, with NVIDIA drivers version 580 or newer recommended for optimal performance. The model download size ranges from 40 to 70 GB, and the engine supports OpenAI and Anthropic API compatibility for local agent integration.written, reviewed, cited
  • RX 6000 cards lack a dedicated installer and must fetch the engine automatically after installing hip-rdna3 or amd-rx7900.
  • Strata never splits a single model across NVIDIA and AMD graphics cards, even if multiple cards of the same make are present.
  • The Coder model requires 32 GB of RAM and retains only half of its experts, making it weaker for non-code tasks.
Qwen 3.8 27B is excellent, but it defaults to overthinking things
simonwillison.net2026-08-16authorityHN front pageabout the Qwen familyHN 802 pts · 385 comments2,116 words
Qwen 3.8 27B is an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba that defaults to an extra high reasoning effort, causing it to overthink simple tasks. The model is a 17GB file capable of vision tasks like bounding box detection and driving coding agents, though generating a complex pelican SVG takes 21 minutes using 22,276 reasoning tokens. Disabling reasoning reduces generation to 137 seconds, while Multi-Token Prediction boosts performance by around 72% compared to the default GGUF setup.written, reviewed, cited
  • The model defaults to xhigh reasoning effort, which is described as a bad place to start for consumer hardware.
  • Running the model with Multi-Token Prediction via llama-server outperformed the default LM Studio GGUF by around 72%.
  • Dense non-Mixture-of-Experts models require a whole lot of memory bandwidth to perform well on available machines.
Qwen 3.8 27B available on Cerebras at 1500 tokens/s
inference-docs.cerebras.ai2026-09-03HN front pageabout the Qwen familyHN 691 pts · 228 comments392 words
Qwen 3.6 27B is the sweet spot for local development
quesma.com2026-06-29HN front pageabout the Qwen familyHN 1,192 pts · 759 comments1,257 words

Community1

Qwen 3.8
twitter.com2026-07-19HN front pageabout the Qwen familyHN 961 pts · 731 comments
All coverage tracked for this model (35) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/sgithub.com2026-10-042026-10-14hn
Qwen Image 2.1qwen.ai official2026-09-202026-10-14hn
Qwen 3.8 27B available on Cerebras at 1500 tokens/sinference-docs.cerebras.ai2026-09-032026-10-14hn
Qwen3.8-Flash-Nextqwen.ai official2026-08-262026-10-14hn
Qwen 3.8 27B is excellent, but it defaults to overthinking thingssimonwillison.net2026-08-162026-10-14hn
Qwen3.8-Max: A New Bar for Coding and Coworkqwen.ai official2026-08-032026-10-14hn
Show HN: Qwen Scribe – local transcription and dictation for Apple Silicongithub.com2026-07-292026-10-14hn
Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledgeqwen.ai official2026-07-212026-10-14hn
Qwen 3.8twitter.com2026-07-192026-10-14hn
Qwen 3.6 27B is the sweet spot for local developmentquesma.com2026-06-292026-10-14hn
Local Qwen isn't a worse Opus, it's a different toolblog.alexellis.io2026-06-182026-10-14hn
Qwen-VLA: Vision-Language-Action Modeling Across Tasks, Environments, and Robotsdcard.tw2026-06-012026-10-14hn
Qwen3.7-Max: The Agent Frontierqwen.ai official2026-05-202026-10-14hn
Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Modelqwen.ai official2026-04-222026-10-14hn
Qwen3.6-Max-Preview: Smarter, Sharper, Still Evolvingqwen.ai official2026-04-202026-10-14hn
Qwen3.6-35B-A3B: Agentic coding power, now open to allqwen.ai official2026-04-162026-10-14hn
Show HN: Using RasPi and Qwen 2.5 VL to find out when birds visit my balconyakshay326.com2026-04-042026-10-14hn
Qwen3.6-Plus: Towards real world agentsqwen.ai official2026-04-022026-10-14hn
1M Tokens/s: Scaling Qwen 3.5 27B on 96 B200 GPUs with vLLMmedium.com2026-03-272026-10-14hn
How to run Qwen 3.5 locallyunsloth.ai2026-03-072026-10-14hn
Something is afoot in the land of Qwensimonwillison.net2026-03-042026-10-14hn
Qwen3-Coder-Nextqwen.ai official2026-02-032026-10-14hn
Qwen3-Max-Thinkingqwen.ai official2026-01-262026-10-14hn
Qwen3-TTS family is now open sourced: Voice design, clone, and generationqwen.ai official2026-01-222026-10-14hn
Qwen3-VL-Embedding and Qwen3-VL-Reranker: Next Gen of Multimodal Retrievalqwen.ai official2026-01-082026-10-14hn
Show HN: Easily Run Qwen-3 VL on HPC-AI.comhpc-ai.com2025-09-292026-10-14hn
Qwen3-VLqwen.ai official2025-09-232026-10-14hn
Qwen3-Nextqwen.ai official2025-09-122026-10-14hn
Tiny-Qwen: Minimal PyTorch Re-Implementation of Qwen3 and Qwen2.5-VLgithub.com2025-08-182026-10-14hn
Qwen-Image: Crafting with native text renderingqwenlm.github.io official2025-08-042026-10-14hn
Qwen VLo: From “Understanding” the World to “Depicting” Itqwenlm.github.io official2025-06-272026-10-14hn
Show HN: Qwen-2.5-32B is now the best open source OCR modelgithub.com2025-04-012026-10-14hn
Introducing Qwen-VL (Alibaba LLM)qwenlm.github.io official2024-01-292026-10-14hn
Qwen-VL: A Frontier Large Vision-Language Model with Versatile Abilitiespaperswithcode.com2023-08-252026-10-14hn
Qwen-VL, a Large Vision Language Model by Alibaba Cloudgithub.com2023-08-232026-10-14hn

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Similar models

3 models

Similar models

  • opt-125m
    same task (text-generation)
  • gpt-j-6b
    same task (text-generation), permissive license
  • DialoGPT-medium
    same task (text-generation), permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-14 14:08 UTC · Last verified 2026-10-25

Follow Qwen-VL

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.