How much VRAM for Qwen 3.6 27B? GPUs, Macs and context
How much VRAM for Qwen 3.6 27B, 3.5 27B and 3.8 27B at Q4: the 24 GB and 32 GB cards and Macs that fit, the context each holds, est. speed and 16 GB misses.
Updated · By RigPerDollar
As an Amazon Associate I earn from qualifying purchases. Some links on this page are affiliate links.
Short answer
Qwen 3.6 27B and 3.5 27B need about 19.5 GB of VRAM at Q4, and Qwen 3.8 27B needs 19 GB. A 24 GB RTX 3090, RTX 4090 or RX 7900 XTX runs all three with 74K to 82K tokens of context, and a 32 GB RTX 5090 holds 194K to 202K. Cards with 16 GB fall 3 to 3.5 GB short.
A few terms
- VRAM is the memory on your graphics card. The model file and its working memory both have to fit there. A Mac or a DGX Spark has one pool of unified memory instead, and a Mac lets its GPU use only part of it by default. By our estimate, a 36 GB Mac gives the GPU about 24.1 GB.
- Q4 is a compressed file format that stores each weight in about 4 bits, so a model fits in less memory. Every size in this guide is a Q4_K_M GGUF file, the format llama.cpp uses.
- Context is how much text the model can hold at once, counted in tokens (pieces of words). Runtimes count it in binary thousands, so 32K is 32,768 tokens and 256K is 262,144.
- The KV cache is the model’s notes on every token it has read. It grows with every token of context. An 8-bit (q8_0) cache stores those notes in about half the space of the default FP16 cache, so more context fits.
- Tok/s means tokens per second, or how fast the model writes. Every speed in our tables is a RigPerDollar estimate for llama.cpp at short context, with a target of about 15% either way.
We rank hardware by fit, context, estimated speed and estimated speed per dollar. We don’t test how good a model’s answers are.
The Qwen 27B versions on the site
Our data has three Qwen 27B models. All three have 27.78B parameters, a 256K context window and an Apache 2.0 license. None of them is a mixture-of-experts (MoE) model in our data, so every token reads the whole file, which makes them slower than MoE models of a similar size.
| Model | Released | Q4 file | Memory needed by our fit rule | Max context | License |
|---|---|---|---|---|---|
| Qwen3.5-27B | Feb 24, 2026 | 17.98 GB | 19.5 GB | 256K | Apache 2.0 |
| Qwen3.6-27B | Apr 21, 2026 | 17.98 GB | 19.5 GB | 256K | Apache 2.0 |
| Qwen3.8-27B | Aug 5, 2026 | 17.44 GB | 19 GB | 256K | Apache 2.0 |
File sizes are the Q4_K_M files from bartowski’s GGUF uploads on Hugging Face, as listed on each model page. By our fit rule, a model fits when its file plus 1.5 GB of headroom fits in memory, and at least 4K tokens of context with an 8-bit cache still fit after that. Context numbers assume a 1 GB runtime reserve. To get Qwen3.8-27B from a source other than Hugging Face, see our Hugging Face alternatives guide.
The three versions share the same KV cache layout in each model’s config.json (3.5, 3.6, 3.8). Each is a hybrid design in which 16 layers keep a KV cache, with 4 KV heads of size 256. That costs 65,536 bytes per token with an FP16 cache and 34,816 bytes with an 8-bit cache. In practice:
| Context | KV cache, FP16 | KV cache, 8-bit |
|---|---|---|
| 4K | 0.27 GB | 0.14 GB |
| 32K | 2.15 GB | 1.14 GB |
| 128K | 8.59 GB | 4.56 GB |
| 256K (full) | 17.18 GB | 9.13 GB |
So the gap between the versions comes from file size alone. Qwen3.8-27B’s file is 0.54 GB smaller, which leaves room for about 8K more tokens of FP16 context on the same card.
Graphics cards that fit Qwen 27B
The first context number in each cell uses the default FP16 cache. The second uses an 8-bit cache. Qwen3.5-27B and Qwen3.6-27B give the same numbers, so they share a column.
| Card | VRAM | Context, 3.5 or 3.6 (FP16 / 8-bit) | Context, 3.8 (FP16 / 8-bit) | Est. tok/s, 3.5 or 3.6 | Est. tok/s, 3.8 |
|---|---|---|---|---|---|
| Radeon RX 7900 XT | 20 GB | 15K / 28K | 23K / 43K | 32 | 33 |
| RTX 3090 | 24 GB | 74K / 140K | 82K / 155K | 39 | 40 |
| Radeon RX 7900 XTX | 24 GB | 74K / 140K | 82K / 155K | 37 | 38 |
| RTX 4090 | 24 GB | 74K / 140K | 82K / 155K | 42 | 43 |
| Radeon AI PRO R9700 | 32 GB | 194K / full 256K | 202K / full 256K | 26 | 26 |
| RTX 5090 | 32 GB | 194K / full 256K | 202K / full 256K | 67 | 68 |
Some notes on the table:
- A 20 GB card is the smallest that fits. It leaves room for less than 32K of context on 3.5 and 3.6 even with an 8-bit cache. Qwen3.8-27B reaches 32K on 20 GB only with the 8-bit cache.
- On any 24 GB card, all three hold 32K with the default cache. For 128K you need the 8-bit cache.
- A 32 GB card holds the full 256K window with an 8-bit cache.
- The RX 7900 XTX has a little more bandwidth than the RTX 3090 (960 GB/s against 936 GB/s), yet it estimates slightly slower. In the published llama.cpp results our estimates are fitted to, AMD cards turn less of their bandwidth into speed (0.77 of bandwidth, against 0.88 for NVIDIA with CUDA).
- The R9700 has 32 GB but 640 GB/s of bandwidth, so it has the room of a 5090 at the same estimated speed as the 16 GB RX 9070 (26 tok/s).
Macs and unified-memory systems
On a Mac, the memory column shows the share the GPU can use by default. A DGX Spark can use all of its memory.
| System | Memory (GPU share) | Context, 3.5 or 3.6 (FP16 / 8-bit) | Context, 3.8 (FP16 / 8-bit) | Est. tok/s, 3.5 or 3.6 | Est. tok/s, 3.8 |
|---|---|---|---|---|---|
| Apple M1 Max, 32 GB | 32 GB (21.4 GB) | 36K / 67K | 44K / 83K | 18 | 19 |
| Mac Studio M5 Max, 32-core GPU | 36 GB (24.1 GB) | 76K / 143K | 84K / 158K | 21 | 21 |
| Mac Studio M5 Max, 40-core GPU | 64 GB (48 GB) | full 256K | full 256K | 27 | 28 |
| Mac Studio M5 Ultra | 96 GB (72 GB) | full 256K | full 256K | 47 | 48 |
| NVIDIA DGX Spark | 128 GB | full 256K | full 256K | 13 | 13 |
A Mac with 32 GB is the smallest that fits by our rule. Any Mac with 24 GB gives the GPU about 16.1 GB, which is 3.4 GB short for 3.5 and 3.6 and 2.9 GB short for 3.8.
Popular cards that miss
These 16 GB cards don’t fit any of the three under the fit rule. The last two columns show how far short they fall.
| Card | VRAM | Short by, 3.5 or 3.6 | Short by, 3.8 |
|---|---|---|---|
| RTX 5060 Ti 16GB | 16 GB | 3.5 GB | 3 GB |
| RTX 5070 Ti | 16 GB | 3.5 GB | 3 GB |
| RTX 5080 | 16 GB | 3.5 GB | 3 GB |
| Radeon RX 9070 | 16 GB | 3.5 GB | 3 GB |
| Radeon RX 9070 XT | 16 GB | 3.5 GB | 3 GB |
You can push part of a model into system RAM, and llama.cpp will run the rest on the CPU. That runs much slower than any number in this guide. My Rig can check whether a model fits your card plus your RAM once you enter your RAM size, and the methodology explains how it treats offload speed.
Two 16 GB cards give 32 GB, and llama.cpp can split the model by layers across them. By our rule, two RTX 5060 Ti 16GB cards hold Qwen3.8-27B with 202K of FP16 context. The split adds memory. The speed estimate stays at one card’s 21 tok/s. Check that your board has room for both cards and that your power supply can feed them.
Qwen also has an MoE sibling on the site. In an MoE model only a few parts, called experts, run for each token. Qwen3.6-35B-A3B and Qwen3.5-35B-A3B have 35.95B parameters, with 3B active per token, and a 22.29 GB Q4 file. They miss 16 GB cards by 7.8 GB. On an RTX 3090 they fit with 33K / 63K of context at an estimated 157 tok/s, and an RTX 5090 holds the full 256K. Our RTX 5070 Ti vs RTX 3090 guide compares the two cards on more models.
Which card per dollar
This table uses Qwen3.8-27B as the test model. Card prices are for the card alone, so you still need a PC to put it in. Mac prices are for a whole computer, so they sit in their own rows below.
| Hardware | Price | Price date and source | Est. tok/s, Qwen3.8-27B | $ per est. tok/s |
|---|---|---|---|---|
| Radeon RX 7900 XTX | $1,350 new | Nov 2, 2026, lowest in-stock Newegg listing | 38 | $35.53 |
| RTX 3090 | $1,499 used | Nov 2, 2026, eBay sold prices | 40 | $37.48 |
| RTX 4090 | $2,910 used | Nov 2, 2026, eBay sold prices | 43 | $67.67 |
| RTX 5090 | $6,500 new | Nov 2, 2026, lowest in-stock Newegg listing | 68 | $95.59 |
| Mac Studio M5 Max, 36 GB (whole computer) | $2,499 new | Oct 16, 2026, apple.com starting price | 21 | $119.00 |
| Mac Studio M5 Ultra, 96 GB (whole computer) | $5,499 new | Oct 23, 2026, apple.com starting price | 48 | $114.56 |
The new card prices come from picksbycard’s daily read of Newegg listings. The used prices come from gpusold.com’s record of completed US eBay sales over 14 days, price plus shipping. For the RTX 3090 that was 295 sales, with $1,499 at the bottom of the typical range ($1,499 to $1,625) and a median of $1,551. For the RTX 4090 it was 103 sales, with a typical range of $2,910 to $3,236 and a median of $3,100.
Among the priced cards, the RX 7900 XTX and the used RTX 3090 give the most estimated speed per dollar, within about $2 of each other. The RTX 5090 costs more per token but is the only priced card that holds the full 256K window. We have no price for the RX 7900 XT, the Radeon AI PRO R9700 or the 64 GB Mac Studio M5 Max.
To see every model’s fit, context and estimated speed on one card, open My Rig set to an RTX 3090. For a Mac, try My Rig set to the 36 GB Mac Studio M5 Max.
Measured runs from other people
These are third-party and self-reported results, kept apart from our estimates above. Each used its own file, runtime and settings, so don’t compare one with another or with our estimates.
- RTX 3090, Qwen3.6-27B Q4_K_M (a 16.80 GB file), llama.cpp build b8777 with CUDA, llama-bench, published by ahelpme.com and dated June 10, 2026 in our records: 40.14 tok/s on one card, and 40.61 tok/s on two cards with a layer split. The same source measured 31.8 tok/s for the Q6_K file on one card. These runs are among the published results our estimates are fitted to.
- MacBook Pro with M5 Max (40-core GPU), Qwen3.6-27B UD-Q4_K_XL (a 17.61 GB file), llama.cpp build b9150 with Metal and no flash attention, posted on note.com and dated May 16, 2026 in our records: 25.6 tok/s.
- One RTX 5090, kaitchup/Qwen3.8-27B-SLQ-EAR98, a different quantized build with 23.73 GB of weight files, run in vLLM 0.31.0 by Niklas Lenz on Oct 25, 2026. The largest context that loaded was 74,752 tokens, per the author. Reported writing speed was 67.2 tok/s for one request. Self-reported, checked by us against the author’s raw output file, not reproduced.
- Two RTX 5060 Ti 16GB cards, the same SLQ-EAR98 build and vLLM version, split with tensor parallel 2, by the same author on Oct 25, 2026, at 94,208 tokens of context: 33.0 tok/s for both cards together. Self-reported, checked by us against the raw file, not reproduced.
What we haven’t verified
- Every tok/s figure in our tables is an estimate. We have not run these models on these cards ourselves.
- We don’t test answer quality, so this guide doesn’t say which of the three versions gives better answers.
- Context limits come from our memory math and config.json data. Real runtimes add their own overhead, which can move the limit.
- We have no price for the RX 7900 XT, the Radeon AI PRO R9700 or the 64 GB Mac Studio M5 Max.
- The used prices are a snapshot of completed eBay sales, and they change week to week.
- RAM offload speed for these models. We have too few measured offload runs to estimate it.