Qwen Open Weights and Local AI — What the Catalog Verifies Today

"Open weights soon" and "you can run this locally today" are still not the same sentence — but Qwen's August 2026 lineup now has three catalog identities, not one promise. The Qwen3.8-Max model profile is the hosted API (openWeights: false): multimodal input, million-token context, and managed tools. The downloadable Max-class checkpoint is Qwen3.8-2.4T-A95B under the Qwen3.8-Max License. The practical local VLM is Qwen3.8-27B under Apache 2.0. Treat those as separate products.

Read the catalog fields, not the headline

Every model profile on this site carries openWeights, verifiedAt, and explicit watch-outs rather than implied promises.

• Hosted Qwen3.8-Max: API status catalogued; openWeights: false. Vision, non-thinking, 1M context, and built-in tools live here.
• Qwen3.8-2.4T-A95B: openWeights: true on Hugging Face. Text-only; thinking cannot be disabled; native context is 262,144 tokens (YaRN toward ~1,010,000). Qwen3.8-Max License — MaaS / AI Work Assistant businesses over $50M trailing 12-month revenue need a separate commercial licence.
• Qwen3.8-27B: openWeights: true, Apache 2.0, dense 27B VLM with image and video input. Thinking is on by default and can be disabled. Hosted 1M-context API is documented as coming soon.

That pattern repeats across the open-weight library. Models like DeepSeek V4-Flash and Kimi K3 carry openWeights: true with licence and scale caveats on their profiles. The DeepSeek V4-Flash open-weights guide and Kimi K3 open-weights guide document 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 weights available under a licence I can use? Check the model profile and the matching open-weights page. 27B is Apache 2.0; 2.4T-A95B is not.
2. Can my hardware serve the model at useful concurrency? The GPUs index and individual GPU profiles (for example RTX 5090 or RTX 4090) document VRAM and editorial fit scores — not guaranteed tokens per second. 2.4T is a datacentre checkpoint; 27B is the first realistic local Qwen3.8 target.
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 models you can self-host today

If your goal is local or provider-flexible deployment now, start with models the catalog already marks as open-weight:

• Qwen3.8-27B — Apache 2.0 dense VLM; see the open-weights guide.
• Qwen3.8-2.4T-A95B — Max-class MoE under the Qwen3.8-Max License; see the open-weights guide.
• DeepSeek V4-Flash — budget-tier API with MIT weights; see the open-weights guide.
• Kimi K3 — open weights with a datacentre-scale footprint; the profile watch-out is explicit about what "open" does not mean for a workstation.
• GLM 5.2 — million-token context with open-weight positioning; GLM 5.3 is the current Coding Plan default and is not open-weight yet.

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 verification workflow that avoids stale assumptions

1. Open the open-weights index and filter to models whose verifiedAt dates you accept.
2. Read the watch-out on each profile — scale, licence, and quantisation caveats live there.
3. If hardware is in scope, cross-check GPU profiles for VRAM fit; if rental is in scope, compare inference providers on editorial planning scores.
4. Re-check after major provider announcements; run verifiedAt on the page, not memory.

The best AI for coding guide reminds teams to measure success on their repository — compile, test, review, rollback — rather than on a launch-week benchmark table.

When a new Qwen identity lands

A new hosted API, a new checkpoint, or a licence change should get its own catalog slug. Do not flip openWeights on Qwen3.8-Max just because a related checkpoint exists. Until a row says openWeights: true with a licence you can live with, do not budget hardware against that name.

A checklist before you buy hardware for "local Qwen"

1. Confirm openWeights: true on the specific profile — Max API, 2.4T-A95B, or 27B.
2. Read the licence on the matching open-weights guide; API terms and weight terms differ.
3. Estimate VRAM from parameter and quantisation notes on the profile; cross-check GPU profiles for fit scores, not forum guesses.
4. If step three fails, compare GPU inference rental against QwenCloud API rates on the verified Max profile before capital spend.
5. Re-run acceptance tests after any weight drop — checkpoints change behaviour even when names stay stable.

Skipping step one is how teams end up with a GPU invoice and no checkpoint they are allowed to load. The library is deliberately conservative here: a missing weight file is not a missing marketing page.