Qwen Open Weights and Local AI — Which Qwen 3.8 Checkpoint You Can Actually Run

Current checkpoint check: 1 October 2026. Model-card facts were checked against the official Qwen Hugging Face cards and the site's dated catalog sources below. Hardware fit is an estimate, not a fit guarantee; validate your runtime, quantisation and context on the target machine.

"Open weights" and "you can run this locally" are still not the same sentence. Qwen's 3.8 generation makes the gap easy to see. There is a hosted flagship API, a datacentre-scale open checkpoint of that flagship, an experimental Qwen 4 preview, and a compact dense model that fits a single high-memory GPU. They have different licences and different hardware needs. Treat each one as a separate product.

The four Qwen 3.8 identities

Two corrections to earlier versions of this post:

• Qwen3.8-27B is downloadable under Apache 2.0. Its exact checkpoint release day is not verified; repository creation is not release proof. It is smaller than the 2.4T checkpoint, and fit depends on quantisation and runtime. The current managed QwenCloud card and Model Studio pricing now list its hosted SKU separately; the checkpoint card's “coming soon” text is not the current hosted availability source.
• The open checkpoints have different licence conditions. The 2.4T-A95B model's Qwen3.8-Max License requires the model name to be displayed when a commercial product or service exceeds 100 million monthly active users or US$20 million in monthly revenue. If a business or affiliate offers MaaS or an AI Work Assistant and its aggregate revenue exceeds US$50 million in any consecutive 12 months, it must obtain a separate licence for commercial use. The Flash Next Qwen Community License 1.0 has the same display threshold but requires a separate licence for commercial MaaS or AI Work Assistant use without that revenue threshold. These are summaries, not legal advice; read each checkpoint's current licence file. The open checkpoint is text-only with 262,144 native context and configured extension toward 1M; the managed API has a different visual-input and default-context contract.

Read the catalog fields, not the headline

Every model profile on this site carries openWeights, license, verifiedAt, and explicit watch-outs rather than implied promises. For Qwen 3.8 that means:

• Qwen 3.8 Max: the catalog profile describes the managed API and has openWeights: false. The separate Qwen3.8-2.4T-A95B checkpoint is downloadable under the Qwen3.8-Max License; its weight licence does not apply to the hosted profile.
• Qwen 3.8 Flash Next: openWeights: true, Qwen Community License 1.0. The official model card describes an experimental architecture preview; its separate licence file has distinct commercial-use conditions from the Max checkpoint.
• Qwen 3.8 27B: openWeights: true, Apache 2.0. Model Studio International now lists $0.50 input/$3.00 output per million tokens, with separate China and cache prices. These are hosted prices, not self-hosting costs; the managed 1M context does not replace the checkpoint's native 262,144 limit.

The same pattern repeats across the open-weight library. Kimi K3 and GLM 5.3 carry licence and scale caveats on their profiles. The Kimi K3 open-weights guide documents self-hosting trade-offs without inventing throughput numbers.

Local AI is three decisions, not one

Teams asking about "running Qwen locally" are usually conflating three questions:

1. Are the weights available under a licence I can use? Apache 2.0 (27B) and a custom Max licence (2.4T-A95B) carry different obligations. Check the exact checkpoint's licence file and get legal advice for commercial use.
2. Can my hardware serve the model at useful concurrency? A simple weight-size estimate for 27B parameters is about 13.5GB at 4-bit and 27GB at 8-bit before quantisation metadata, runtime overhead and KV cache; it is not a VRAM-fit guarantee. The 2.4T-parameter checkpoint is a multi-GPU/datacentre-scale deployment. The GPUs index and Local AI Model Fit Finder are starting points; verify the exact quant and context on target hardware.
3. Is self-hosting cheaper than a hosted API at my measured throughput? Compare against GPU inference providers and managed APIs before buying silicon.

The library's replace closed API with open weights switch guide treats this as a migration with a pilot plan, not a weekend project.

Qwen vs other models you can self-host today

• Qwen 3.8 27B: Apache 2.0, dense and multimodal. It is a smaller local option; check the hardware fit for your chosen quant and context.
• Muse Glimmer 30B: Meta's open-weight 30B model. Meta says it is designed to run on a single GPU or a Mac Mini, which makes it the obvious side-by-side test against Qwen3.8-27B.
• Qwen3.8-2.4T-A95B: the open 2.4T/95B checkpoint for teams with cluster capacity. Remember the custom licence and the capabilities the open build drops.
• Qwen3-235B-A22B: the previous-generation Apache 2.0 MoE, still a good multi-GPU option.
• Kimi K3: open weights with a datacentre-scale footprint. The profile is explicit about what "open" does not mean for a workstation.
• GLM 5.3: a 753B-total MoE under a custom licence; its active count is not verified. GLM 5.3 Flash is the separate MIT-licensed 320B/18B model.

For a hosted multimodal frontier path, the Qwen3.8-Max vs GPT-5.6 Sol comparison and the switch guide frame parallel evaluation, not rip-and-replace.

A checklist before you buy hardware for "local Qwen"

1. Pick the specific checkpoint: 27B, Flash Next or 2.4T-A95B. Do not budget hardware against the "Qwen 3.8" family name.
2. Read the licence on the Hugging Face model card. API terms and weight terms differ, and the Max checkpoint's licence is custom.
3. Estimate VRAM from parameter count and quantisation, then cross-check GPU profiles for fit scores rather than forum guesses.
4. If step three fails, compare GPU inference rental against the hosted Max API before spending capital.
5. Re-run acceptance tests after any new checkpoint. Behaviour changes even when names stay stable.

Skipping step one is how teams end up with a GPU invoice and a checkpoint they are not allowed to use commercially. When a new Qwen checkpoint or licence change lands, it should get its own catalog row. Do not assume it inherits the terms of a sibling.

Sources

• Hugging Face — Qwen/Qwen3.8-27B model card (27B dense vision-language model; published 2026-08-13; Apache 2.0; 262K native context): https://huggingface.co/Qwen/Qwen3.8-27B — accessed 2026-10-01
• Hugging Face — official Qwen/Qwen3.8-2.4T-A95B model card and its LICENSE file (2.4T total / 95B active; exact Qwen3.8-Max License conditions): https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B and https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/resolve/main/LICENSE — accessed 2026-10-01
• Hugging Face — official Qwen/Qwen3.8-Flash-Next model card and its LICENSE file (experimental Qwen4 architecture preview; 125B main model, 6B active, 51B n-gram embeddings, 4B MTP; exact Qwen Community License 1.0 conditions): https://huggingface.co/Qwen/Qwen3.8-Flash-Next and https://huggingface.co/Qwen/Qwen3.8-Flash-Next/resolve/main/LICENSE — accessed 2026-10-01
• Hugging Face — unsloth/Qwen3.8-2.4T-A95B-GGUF (2.4T total / 95B active MoE; community GGUF of the open Max checkpoint): https://huggingface.co/unsloth/Qwen3.8-2.4T-A95B-GGUF — accessed 2026-09-26
• Qwen — Qwen3.8-Flash-Next announcement: https://qwen.ai/blog?id=qwen3.8-flash-next — cited via the site catalog row (verified 2026-09-03)
• Meta — Meta Connect 2026 recap (Muse Glimmer 30B open weights, "single GPU or Mac Mini"): https://developers.meta.com/blog/meta-connect-recap/ — accessed 2026-09-26
• Qwen — model organisation on Hugging Face (Qwen3-235B-A22B, Apache 2.0): https://huggingface.co/Qwen — accessed 2026-08-28