Apache-2.0multimodalRuns locally27B (card)
Qwen3.8-27B: Specs, License & How to Run
Apache-2.0 27B native vision-language model from Qwen with 262K context.
- License
- Apache-2.0
- Parameters
- 27B (card)
- HF likes
- 18,393 (2026-10-26)
- First seen
- 2026-08-05
TL;DR. Qwen3.8-27B is a multimodal model from Qwen. Parameters: 27B (card).License: Apache-2.0. It is listed under ChatGPT, Closed LLM APIs.
| Developer | Qwen |
|---|---|
| Type | multimodal |
| Architecture | Causal language model with a vision encoder (HF model_type qwen3_5; architecture Qwen3_5ForConditionalGeneration). Card: 64 layers, hidden 5120, 24 Q / 4 KV attention heads, FFN 17408, vocab 248320 padded, hybrid Gated DeltaNet + Gated Attention. pipeline_tag image-text-to-text. Native image and video understanding. |
| Parameters | 27B (card) |
| Modalities | text, image, video |
| Context | 262,144 tokens native (card + text_config); extensible to 1,000,000 (card) |
| License | Apache-2.0 — LICENSE on the Hugging Face card is Apache License 2.0. cardData.license is apache-2.0 (opened 2026-10-26). Hosted Qwen Cloud terms are separate from the weight license. Not legal advice. |
| Hardware | Not yet verified — HF usedStorage is 55,615,875,685 bytes (~51.8 GiB). safetensors.parameters sum to 27,781,427,952. Peak VRAM is not independently measured. The card names SGLang, vLLM, and TokenSpeed; hosted Qwen Cloud is a separate product. (not independently tested) |
| Hugging Face likes | 18,393 (as of 2026-10-26) |
| Model card | https://huggingface.co/Qwen/Qwen3.8-27B |
| First seen | 2026-08-05 |
Qwen3.8-27B is Qwen’s ungated 27B native vision-language checkpoint. Official hosts we opened on 2026-10-26: the card Qwen/Qwen3.8-27B (created 2026-08-05T08:22:59Z; last modified 2026-08-14T15:00:01Z; 18,393 likes / 7,762,014 downloads), the Apache 2.0 LICENSE, qwen.ai, and the named overview qwencloud.com/models/qwen3.8-27b. API gated was false.
This page hangs off ChatGPT alternatives and closed LLM APIs. That mapping is editorial. We have no sourced quality comparison against ChatGPT. Same-site decision post-train of this checkpoint: Cloudflare Clef. Head-to-head facts versus the 33B teacher that names this card: Nanbeige4.3-33B vs Qwen3.8-27B.
The README calls Qwen3.8 “the most capable generation in the Qwen open-model family to date” and describes coding, professional work, research, long-horizon agents, and native image/video understanding. Those adjectives stay vendor. Hugging Face pipeline_tag is image-text-to-text; tags include qwen3_5. Thinking mode is on by default; the card names reasoning_effort and preserve_thinking.
What we verified (2026-10-26)
| Field | Value | Source |
|---|---|---|
| Host | huggingface.co/Qwen/Qwen3.8-27B (not gated) | HF API |
| Created / last modified | 2026-08-05T08:22:59Z / 2026-08-14T15:00:01Z | same |
| Likes / downloads | 18,393 / 7,762,014 | same |
| License tag + file | apache-2.0 / Apache License 2.0 | HF cardData + LICENSE |
| Pipeline tag | image-text-to-text | HF API |
| Architecture | qwen3_5 / Qwen3_5ForConditionalGeneration | config.json |
| Params (card) | 27B | card |
| safetensors.parameters | 27,781,427,952 | HF API |
| usedStorage (BF16) | 55,615,875,685 bytes (~51.8 GiB) | same |
| Layers / hidden / heads | 64 / 5120 / 24 Q, 4 KV | card + text_config |
| Native context | 262,144 tokens | card + max_position_embeddings |
| Stated extension | 1,000,000 tokens | card |
| Job | Images + video + text in, text out | card |
How to run (from the card, not tested)
The card names SGLang, vLLM, and TokenSpeed, and points at Qwen3.8 cookbooks rather than pasting a single flag set on the README we opened. Thinking-mode sampling copied from the card: temperature 1.0, top_p 0.95, top_k 20. Instruct / non-thinking settings are also listed on that README. forkalt has not loaded the checkpoint.
Hosted Qwen Cloud is a different artifact from the Apache-2.0 weights. The card said a hosted 27B with 1M context and built-in tools was coming; treat that page’s current status as upstream.
License
LICENSE is Apache License 2.0. Card tag apache-2.0. Hosted API terms are separate. Not legal advice.
What we did not invent
- Quality adjectives and any Qwen-published leaderboard rows stay vendor.
- We do not copy Nanbeige’s self-reported opponent column onto this page as a Qwen claim.
- Peak VRAM is not on the card extract we used.
- We do not rank Qwen3.8-27B against ChatGPT, Nanbeige4.3-33B, or Cloudflare Clef.
Related: Nanbeige4.3-33B — Nanbeige4.3-33B vs Qwen3.8-27B — Cloudflare Clef (27B decision post-train) — SenseNova 6.8 Code2Video (card: Qwen3.8-27B fine-tune for code-driven video; different job) — Flint, Sage-A1.5. Serving: vLLM.
Deploy options
Comparisons
FAQ
Is Qwen3.8-27B open weight and Apache-2.0?
The ungated card Qwen/Qwen3.8-27B (created 2026-08-05T08:22:59Z; 18,393 likes / 7,762,014 downloads in the 2026-10-26 API check) tags license apache-2.0. The LICENSE file is Apache License 2.0. Hosted Qwen Cloud is not the weight grant. Not legal advice.
Is this text-only?
No. Hugging Face pipeline_tag is image-text-to-text. The card describes a native vision-language model for images and videos. text_config max_position_embeddings is 262,144; the README also names extension to 1,000,000 tokens.
Why is this page on forkalt now?
Nanbeige4.3-33B’s official table names Qwen3.8-27B as the lead ~30B opponent. This is an identification page so that comparison can cite a same-site card. Cloudflare Clef is already documented as a Qwen3.8-27B post-train.
Does it replace ChatGPT?
No sourced bake-off. Capability adjectives on the Qwen card stay vendor. See ChatGPT alternatives and Nanbeige4.3-33B vs Qwen3.8-27B.
Sources & verification
- https://huggingface.co/Qwen/Qwen3.8-27B
- https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/LICENSE
- https://qwen.ai
- https://www.qwencloud.com/models/qwen3.8-27b
- https://huggingface.co/Cloudflare/clef
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.