Qwen3.8-27B — the 24 GB pick changed hands, and we were twelve days late noticing

Qwen3.8-27B landed on August 5 with 267,000 downloads and 10,400 likes in its first twelve days, 480 community quantizations, and an Apache 2.0 licence. It is the direct successor to the model our planner has recommended at the 24 GB tier since April. We did not catch it in the August 13 sweep, which is worth saying plainly before anything else.

Verdict: The most-downloaded open-weight model of the month, Apache 2.0, and it halves the KV cache at the 24 GB tier

The take

Why it was missed is instructive. It dropped three days after Alibaba opened the 2.4-trillion-parameter Qwen3.8-Max, which is the release everyone wrote about, and our sweep followed the headlines — Muse Glimmer, DeepSeek V4-Pro, Qwen3.8-Max. The quiet 27B that actually runs on readers' hardware went past us while we covered three models almost nobody can run locally. A sweep that tracks announcements rather than adoption will keep making this mistake.

What it is, from the model card: 27.8B dense parameters, 64 layers, a **native vision encoder** (images and video in, not a text model with an adapter attached), 262,144 native context extensible to 1,000,000, and multi-token prediction trained in — so speculative decoding works without a separate drafter. Apache 2.0. Roughly 17 GB at Q4, the same envelope as the Qwen 3.6-27B it replaces.

The architecture is the real change, and it is not cosmetic. The layout is 16 × (3 × Gated DeltaNet → 1 × Gated Attention) — the same linear-attention family Alibaba used in Qwen3.8-Max. Only 16 of the 64 layers keep a KV cache that grows with context; the other 48 hold a constant-size state. Working the published head configuration through (16 layers × K and V × 4 KV heads × 256 head-dim × 2 bytes) gives about 64 KB per token, or **~2.0 GB at 32K context — half the ~4.0 GB the dense 3.6-27B needs at the same parameter count.** On a 24 GB card that difference is the difference between a comfortable long-document session and swapping to disk, and it is why the fit math on our model pages now shows this model reaching context lengths its predecessor could not.

Our call: **Qwen3.8-27B displaces Qwen 3.6-27B as the first pick at `chat.top` and `docs.top`**, with a new entry at `/models/qwen-3-8-27b/`; the 3.6 entry stays as the dated record. Note carefully what that decision does *not* rest on. The benchmark tables on the card are vendor-run, evaluated with Alibaba's own harness, and no independent numbers exist yet — under our own rule those do not displace an incumbent. What justifies the swap is that this is the same vendor's direct successor at the same size, same licence and same hardware target, with roughly two orders of magnitude more adoption than anything else released this month and a structurally cheaper long-context profile we can verify by arithmetic rather than take on trust.

Runner situation is unusually good for a fresh Qwen release: the native Ollama tag `qwen3.8:27b` already resolves, where Qwen models have historically taken one to two weeks, and `unsloth/Qwen3.8-27B-GGUF` is at about 1.9 million downloads. That breadth is worth more than convenience. After Microsoft withdrew Lens in August — a model we had recommended for ten weeks, whose weights simply stopped being downloadable — we said we would prefer picks that live in many community repos over ones that live only in a vendor org. A model with 480 published quantizations is very hard to make disappear.

Where this fits

Models: Qwen3.8-27B · Qwen 3.6-27B · Qwen 3.6-35B-A3B · Qwen3.8-Max (2.4T-A95B)

Hardware: NVIDIA RTX 5090 · NVIDIA RTX 4090 · NVIDIA RTX 3090 (used, single) · Mac Studio M4 Max 64 GB

Sources

Next step

Try this in the planner→