Qwen3-235B-A22B: open-weights frontier MoE, Quarantine

Alibaba's Qwen3-235B-A22B is the flagship of the Qwen3 family: a 235-billion-parameter mixture-of-experts that activates ~22B parameters per token, released under Apache 2.0. It carries the Quarantine tier in our registry. This post sticks to specs we can source — for benchmark numbers we link to Qwen's own report rather than reprint figures we didn't run.

Updated August 11, 2026 — this post originally reported the Sealed tier. Sealed depended on external witness cosigns that were never genuinely third-party, so the tier was retired and no artifact holds it. The registry tier stated here is the one it actually carries.

tldr

235B total / ~22B active MoE (128 experts, 8 routed per token). Apache 2.0 — genuinely permissive. Hybrid "thinking / non-thinking" modes toggled by prompt. 32K native context, up to 128K with YaRN. Released Apr 29, 2025. It's a datacenter-class model — FP8 weights are ~220 GB — not an edge model. Verify its tier and provenance on the live artifact page.

parameters
235 B total
active / token
~22 B (8 of 128 experts)
context
32K native · 128K (YaRN)
architecture
MoE transformer
license
Apache 2.0
on-disk
~220 GB FP8 · ~438 GB BF16
tier (registry)
QUARANTINE
released
2025-04-29

What it is

Qwen3-235B-A22B is a sparse mixture-of-experts model: of its 235B total parameters, a router activates 8 of 128 experts per token, so roughly 22B parameters do the work on any given forward pass. That's the MoE trade — frontier-scale capacity at a fraction of the per-token compute of an equivalently-capable dense model. The original April 2025 release shipped with Qwen3's defining feature: a single checkpoint that can run in a deliberate "thinking" mode or a faster "non-thinking" mode, switched by a prompt directive rather than by loading a separate model.

It is licensed Apache 2.0 — one of the genuinely permissive open-weight licenses, with no acceptable-use rider or scale gate. For capability and benchmark figures, see Qwen's own write-up and model card rather than numbers retyped here:

Sizing it honestly

This is a datacenter model, not an edge one. The native FP8 weights are roughly 220 GB on disk; BF16 is roughly 438 GB. Even at FP8 you need multiple high-memory accelerators (e.g. several H100/H200-class GPUs) to hold the weights plus KV cache. Anyone promising a 235B MoE on a single edge device is selling something — the arithmetic doesn't allow it.

Why it's Quarantine

In our registry, Quarantine means we hold the publisher's own hash for these weights, logged and signed — but we have not yet downloaded the 220 GB ourselves and checked that hash against what actually arrives. That last step is what promotes an artifact to Hardened, and it is the only tier where the number you read is one we produced. You don't have to take any of this on faith: open the live artifact page to see its current tier and provenance, or check any page at /verify.

realistic deployment footprint

Precision
Weights
Runtime
Fits from
FP8
native
~220 GB
vLLM / SGLang
multi-GPU
FP8
~220 GB + KV
≥ 4× H100/H200 80GB class
BF16
~438 GB
8× 80GB-class GPUs

more releases