Qwen3.8-27B VRAM Requirements: How Much Hardware You'll Really Need

Around 17 GB of VRAM is the number circulating for Qwen3.8-27B — Unsloth's Daniel Han said it "should fit in roughly 17GB of VRAM on release" — and it is worth being precise about what that is and is not. It is a projection based on the 27B's predecessor, not a spec from Alibaba, because the model has not shipped yet; the official repository was promised for the week of August 10, 2026 and had not appeared at the time of writing. This article sorts the VRAM question into what you can plan on, what is genuinely uncertain, and how the number moves with your quantization, your context window, and your runtime.

The honest answer first

There is no official hardware spec for Qwen3.8-27B yet. Everything below is projected from Qwen3.6-27B, the model the 27B succeeds — same parameter count, same likely hybrid architecture, and the closest thing there is to a sizing reference. Treat the numbers as a planning envelope, not a requirements sheet. The first real measurements will come from quantizers and inference-framework maintainers within days of the weight drop, and those are the figures to lock your purchase to.

The quant ladder

How much VRAM you need depends almost entirely on which quantization you run. Working up from the community-measured Qwen3.6-27B table:

• Q4_K_M — roughly 16 GB VRAM. The 24 GB-card sweet spot, and the likely home of that "17 GB" figure. Most teams' default.

• Q3_K_M / IQ3 — roughly 13 GB VRAM. Fits 16 GB and even 12 GB cards, with a visible quality step down.

• Q6_K — roughly 21 GB VRAM. Near-lossless, and a tight-but-real fit on a 24 GB card.

• Q8_0 — roughly 27–30 GB VRAM. For 48 GB cards where you want the last drops of quality.

• FP8 serving — roughly 27 GB VRAM for the weights, which is why a single L40S is the common inference-GPU choice.

• Full precision (BF16) — roughly 54 GB VRAM. Realistically an H100-class card, and the territory where people stop calling it "local."

The short version: a 24 GB GPU is the safe purchase for this model, a 16 GB card can run it only if you accept the low quants, and anything with 8 GB or less is out unless the model turns out dramatically leaner than its predecessor.

The variable nobody quotes: context length

Every number above is for a modest context. VRAM scales with context because the KV cache has to live alongside the weights. On Qwen3.6-27B, a 2K context ran in about 11 GB, a 32K context pushed the same quant past 14 GB, and 128K more than doubled the cache footprint. If Qwen3.8-27B keeps a large native context window — and the Qwe​n generation has been trending that way — plan a few GB of headroom beyond the quant size, or cap the context in your runtime config. Picking a context length you do not actually use is the most common way teams end up buying a bigger card than they needed.

The hybrid-architecture wildcard

Qwen3.6-27B was a hybrid model: only 16 of its 65 layers used a traditional KV cache, while 48 used a fixed recurrent state. The result was roughly four times less KV memory than a dense model of the same size — which is a large part of why a 27B felt like a 24 GB card model rather than a 48 GB one. If Qwen3.8-27B keeps the hybrid design (likely, but unconfirmed), the context-length math above gets friendlier than it looks. The catch is runtime support: a llama.cpp or vLLM build that does not implement the recurrent layers will report much higher memory usage. Check which runtimes have merged support before assuming the projections hold.

Which GPUs actually work

Concretely, for the common cards:

• RTX 4090 / 3090 / 5090 (24 GB) — Q4_K_M comfortably, Q6_K if you are willing to squeeze. This is the intended audience.

• RTX 3060 / 4060 Ti 16 GB — IQ4 or Q3-tier quants only, with modest context. Workable, not pleasant.

• 12 GB cards — Q3_K_M at reduced quality. Fine for experimentation, not for serving.

• Apple Silicon — a 24 GB Mac runs the same Q4 files via MLX or llama.cpp; the 16 GB base models swap-thrash and are effectively out, as they were with Qwen3.6-27B.

• 48 GB and up (A6000, L40S, dual-card setups) — Q8_0 or FP8 serving with real headroom.

Or skip the GPU entirely

If the hardware math does not work for you, the model does not have to. OrcaRouter is one API over 200-plus models — the Qwe​n family included — and it passes through provider list price with zero markup, so Qwen3.8-Max currently costs $2.00 per million input tokens and $6.00 per million output tokens, the same as on the vendor's own pricing page. The 27B itself will be a local model, but when its weights drop and it earns trust, the same key will route to whatever provider hosts it — you can build against the generation now and never touch the GPU-reseller market at all.

The bottom line for the hardware question: plan on 16 GB for a comfortable 4-bit run, 24 GB to be safe and to give yourself headroom for context and higher quants, and treat every figure here as provisional until the repository lands and the first real measurements appear. The "17 GB" projection is a sensible planning number — it is just not yet a fact.