Qwen3.6-35B-A3B vs Qwen3.8-27B (Plus Recursive Flint 40B MoE): Which Open Model Fits Your 24 GB or 32–128 GB Rig?
Key facts
Qwen3.8-27B needs 17.80 GB at 4-bit. Qwen3.6-35B-A3B needs 23.10 GB at 4-bit, with 3B active parameters. Our estimate.
Updated . Checked . the home chat.
Both Qwen3.8-27B (Qwen 3.8 27B, 27.78B dense, our estimate) and Qwen3.6-35B-A3B (35.95B Mixture-of-Experts with 3B active parameters per token) fit inside a single 24 GB graphics card (23.4 GB usable VRAM after our 0.6 GB reserve) at standard 4-bit quantization and 8,192 context. Qwen3.8-27B needs 17.80 GB total in our estimate (16.46 GB measured UD-Q4_K_M file + 0.54 GB conversation memory + 0.80 GB working space), leaving 5.60 GB of headroom to step up to 5-bit UD-Q5_K_XL (22.22 GB) or 6-bit UD-Q6_K (23.32 GB) on one 24 GB card. Qwen3.6-35B-A3B needs 23.10 GB total at UD-Q4_K_M (22.13 GB measured file + 0.17 GB conversation memory + 0.80 GB working space, leaving 0.30 GB on 24 GB) or 18.70 GB at compact 4-bit UD-IQ4_XS (17.73 GB file), while activating only 3B parameters per token.
The short answer
Qwen3.8-27B (Qwen 3.8 27B) at UD-Q4_K_M is 17.80 GB in our estimate, UD-Q5_K_XL is 22.22 GB, and UD-Q6_K is 23.32 GB. Qwen3.6-35B-A3B at UD-Q4_K_M is 23.10 GB, at compact UD-IQ4_XS is 18.70 GB, and at UD-Q3_K_XL is 17.82 GB. A 24 GB card leaves 23.4 GB usable. Recursive Flint is 39.79B total with 2.5B active: Q4_K_M is 24.46 GB, which is over 23.4 GB usable, while UD-Q3_K_XL is 19.56 GB and fits a 24 GB card. These totals are our estimate.
Test Recursive Flint free in the browser at the home chat. Qwen3.6-35B-A3B and Qwen3.8-27B are not on that free list. Open Can I run Qwen3.6-35B-A3B?, Can I run Qwen3.8-27B?, and Can I run Recursive Flint?.
Architecture, memory, and licence comparison
| Qwen3.6-35B-A3B | Qwen3.8-27B | Recursive Flint | |
|---|---|---|---|
| Parameters | 35.95B total / 3B active, our estimate | 27.78B dense, our estimate | 39.79B total / 2.5B active |
| Architecture | 40 layers: 10 full attention and 30 linear attention | 64 layers: 16 full attention and 48 linear attention | 49 layers: 14 full attention and 35 sliding-window attention |
| Max context | 262,144 tokens | 262,144 tokens | 262,144 tokens |
| Conversation memory at 8,192 tokens | 0.17 GB | 0.54 GB | 0.27 GB |
| Standard 4-bit file and total | 22.13 GB UD-Q4_K_M, 23.10 GB total | 16.46 GB UD-Q4_K_M, 17.80 GB total | 23.39 GB Q4_K_M, 24.46 GB total |
| Best 24 GB download | UD-Q4_K_M 23.10 GB or UD-IQ4_XS 18.70 GB | UD-Q6_K 23.32 GB or UD-Q5_K_XL 22.22 GB | UD-Q3_K_XL 19.56 GB |
| Best 16 GB download | UD-Q3_K_XL 17.82 GB does not fit 15.4 GB usable | UD-IQ3_XXS 12.27 GB | UD-Q2_K_XL 14.63 GB |
| Licence | Apache-2.0 | Apache-2.0 | Apache-2.0 |
A 16 GB card leaves 15.4 GB usable. Qwen3.8-27B at UD-IQ3_XXS is 12.27 GB (10.93 GB measured file). Recursive Flint at UD-Q2_K_XL is 14.63 GB (13.56 GB measured file). Qwen3.6-35B-A3B's smallest file on our list, UD-Q3_K_XL, is 17.82 GB and does not fit. See what fits in 16 GB and what fits in 24 GB.
Measured downloads across every quality level
Every file size below is a measured download on our list. The total adds conversation memory at 8,192 tokens and working space, and those two parts are our estimate. No Apple MLX file is on our list for these three models.
| Quality | Download | File | Conversation memory | Total | Where it fits | Checked |
|---|---|---|---|---|---|---|
| UD-Q3_K_XL | unsloth/Qwen3.6-35B-A3B-GGUF | 16.85 GB | 0.17 GB | 17.82 GB | 24.0 GB (23.4 GB usable) | checked 24 Oct 2026 |
| UD-IQ4_XS | unsloth/Qwen3.6-35B-A3B-GGUF | 17.73 GB | 0.17 GB | 18.70 GB | 24.0 GB (23.4 GB usable) | checked 24 Oct 2026 |
| MXFP4_MOE | unsloth/Qwen3.6-35B-A3B-GGUF | 21.71 GB | 0.17 GB | 22.68 GB | 24.0 GB (23.4 GB usable) | checked 24 Oct 2026 |
| UD-Q4_K_M | unsloth/Qwen3.6-35B-A3B-GGUF | 22.13 GB | 0.17 GB | 23.10 GB | 24.0 GB (23.4 GB usable) | checked 24 Oct 2026 |
| UD-Q4_K_XL | unsloth/Qwen3.6-35B-A3B-GGUF | 22.36 GB | 0.17 GB | 23.33 GB | 24.0 GB (23.4 GB usable) | checked 24 Oct 2026 |
| Q8_0 | unsloth/Qwen3.6-35B-A3B-GGUF | 36.90 GB | 0.17 GB | 37.87 GB | 64.0 GB (48.0 GB usable) | checked 24 Oct 2026 |
| Quality | Download | File | Conversation memory | Total | Where it fits | Checked |
|---|---|---|---|---|---|---|
| UD-IQ2_XXS | unsloth/Qwen3.8-27B-GGUF | 7.27 GB | 0.54 GB | 8.61 GB | 12.0 GB (11.4 GB usable) | checked 24 Oct 2026 |
| UD-IQ3_XXS | unsloth/Qwen3.8-27B-GGUF | 10.93 GB | 0.54 GB | 12.27 GB | 16.0 GB (15.4 GB usable) | checked 24 Oct 2026 |
| UD-IQ4_XS | unsloth/Qwen3.8-27B-GGUF | 14.25 GB | 0.54 GB | 15.59 GB | 24.0 GB (23.4 GB usable) | checked 24 Oct 2026 |
| UD-Q4_K_M | unsloth/Qwen3.8-27B-GGUF | 16.46 GB | 0.54 GB | 17.80 GB | 24.0 GB (23.4 GB usable) | checked 24 Oct 2026 |
| UD-Q4_K_XL | unsloth/Qwen3.8-27B-GGUF | 17.56 GB | 0.54 GB | 18.90 GB | 24.0 GB (23.4 GB usable) | checked 24 Oct 2026 |
| UD-Q5_K_XL | unsloth/Qwen3.8-27B-GGUF | 20.88 GB | 0.54 GB | 22.22 GB | 24.0 GB (23.4 GB usable) | checked 24 Oct 2026 |
| UD-Q6_K | unsloth/Qwen3.8-27B-GGUF | 21.98 GB | 0.54 GB | 23.32 GB | 24.0 GB (23.4 GB usable) | checked 24 Oct 2026 |
| Q8_0 | unsloth/Qwen3.8-27B-GGUF | 29.05 GB | 0.54 GB | 30.39 GB | 32.0 GB (31.4 GB usable) | checked 24 Oct 2026 |
| Quality | Download | File | Conversation memory | Total | Where it fits | Checked |
|---|---|---|---|---|---|---|
| UD-IQ1_M | unsloth/flint-GGUF | 11.35 GB | 0.27 GB | 12.42 GB | 16.0 GB (15.4 GB usable) | checked 23 Oct 2026 |
| UD-Q2_K_XL | unsloth/flint-GGUF | 13.56 GB | 0.27 GB | 14.63 GB | 16.0 GB (15.4 GB usable) | checked 23 Oct 2026 |
| UD-IQ3_XXS | unsloth/flint-GGUF | 14.54 GB | 0.27 GB | 15.61 GB | 24.0 GB (23.4 GB usable) | checked 23 Oct 2026 |
| UD-Q3_K_XL | unsloth/flint-GGUF | 18.49 GB | 0.27 GB | 19.56 GB | 24.0 GB (23.4 GB usable) | checked 23 Oct 2026 |
| IQ4_XS | bartowski/RecursiveSI_flint-GGUF | 22.45 GB | 0.27 GB | 23.52 GB | 32.0 GB (31.4 GB usable) | checked 23 Oct 2026 |
| NVFP4 | RecursiveSI/flint-nvfp4-gguf | 22.86 GB | 0.27 GB | 23.93 GB | 32.0 GB (31.4 GB usable) | checked 23 Oct 2026 |
| Q4_K_M | ggml-org/Flint-GGUF | 23.39 GB | 0.27 GB | 24.46 GB | 32.0 GB (31.4 GB usable) | checked 23 Oct 2026 |
| UD-Q4_K_XL | unsloth/flint-GGUF | 25.05 GB | 0.27 GB | 26.12 GB | 32.0 GB (31.4 GB usable) | checked 23 Oct 2026 |
| INT4-Q4_1 | RecursiveSI/flint-int4-gguf | 25.99 GB | 0.27 GB | 27.06 GB | 32.0 GB (31.4 GB usable) | checked 23 Oct 2026 |
| Q5_K_M | bartowski/RecursiveSI_flint-GGUF | 30.81 GB | 0.27 GB | 31.88 GB | 64.0 GB (48.0 GB usable) | checked 23 Oct 2026 |
| Q6_K | bartowski/RecursiveSI_flint-GGUF | 35.20 GB | 0.27 GB | 36.27 GB | 64.0 GB (48.0 GB usable) | checked 23 Oct 2026 |
| UD-Q6_K_XL | unsloth/flint-GGUF | 36.32 GB | 0.27 GB | 37.39 GB | 64.0 GB (48.0 GB usable) | checked 23 Oct 2026 |
| Q8_0 | ggml-org/Flint-GGUF | 42.35 GB | 0.27 GB | 43.42 GB | 64.0 GB (48.0 GB usable) | checked 23 Oct 2026 |
| BF16 | ggml-org/Flint-GGUF | 79.63 GB | 0.27 GB | 80.70 GB | 128.0 GB (96.0 GB usable) | checked 23 Oct 2026 |
VRAM tiers: 12 GB, 16 GB, 24 GB, 32 GB, 64 GB, and 128 GB.
Long-context conversation memory (8K to 262K tokens)
Qwen3.6-35B-A3B keeps conversation memory on 10 full-attention layers out of 40 (30 linear-attention layers add none in our estimate). Qwen3.8-27B uses 16 full-attention layers out of 64. That is why the same context costs less VRAM on the mixture-of-experts model. Totals below use the measured 4-bit file plus conversation memory plus 0.80 GB of working space. Our estimate.
| Context tokens | Qwen3.6 conversation memory | Qwen3.6 UD-Q4_K_M total | Qwen3.6 UD-IQ4_XS total | Qwen3.8 conversation memory | Qwen3.8 UD-Q4_K_M total |
|---|---|---|---|---|---|
| 8,192 | 0.17 GB | 23.10 GB at UD-Q4_K_M | 18.70 GB at UD-IQ4_XS | 0.54 GB | 17.80 GB at UD-Q4_K_M |
| 32,768 | 0.67 GB | 23.60 GB at UD-Q4_K_M | 19.20 GB at UD-IQ4_XS | 2.15 GB | 19.41 GB at UD-Q4_K_M |
| 65,536 | 1.34 GB | 24.27 GB at UD-Q4_K_M | 19.87 GB at UD-IQ4_XS | 4.29 GB | 21.55 GB at UD-Q4_K_M |
| 131,072 | 2.68 GB | 25.61 GB at UD-Q4_K_M | 21.21 GB at UD-IQ4_XS | 8.59 GB | 25.85 GB at UD-Q4_K_M |
| 262,144 | 5.37 GB | 28.30 GB at UD-Q4_K_M | 23.90 GB at UD-IQ4_XS | 17.18 GB | 34.44 GB at UD-Q4_K_M |
Speed and community measurements across GPUs, Macs, and mini PCs
Qwen3.8-27B speeds below are our estimate from memory bandwidth, not a measured run. Qwen3.6-35B-A3B and Recursive Flint publish no speed estimate: the estimate would read every weight, which overstates the work a mixture of experts does per token.
| Machine | Usable memory | Price | Qwen3.6-35B-A3B | Qwen3.8-27B | Recursive Flint |
|---|---|---|---|---|---|
| GeForce RTX 3090 24GB | 23.4 GB | €1,246 / $1,395 | no speed estimate (mixture of experts) | about 43 tokens per second, our estimate | no speed estimate (mixture of experts) |
| GeForce RTX 4090 24GB | 23.4 GB | €2,185 / $2,446 | no speed estimate (mixture of experts) | about 46 tokens per second, our estimate | no speed estimate (mixture of experts) |
| GeForce RTX 5090 32GB | 31.4 GB | €7,367 / $6,816 | no speed estimate (mixture of experts) | about 81 tokens per second, our estimate | no speed estimate (mixture of experts) |
| Radeon RX 7900 XTX 24GB | 23.4 GB | €1,254 / $1,404 | no speed estimate (mixture of experts) | about 42 tokens per second, our estimate | no speed estimate (mixture of experts) |
| Intel Arc Pro B70 32GB | 31.4 GB | no euro price on our list | no speed estimate (mixture of experts) | about 17.8 tokens per second, our estimate | no speed estimate (mixture of experts) |
| Radeon RX 9060 XT 16GB | 15.4 GB | €564 / $522 | no speed estimate (mixture of experts) | does not fit | no speed estimate (mixture of experts) |
| Mac mini M4 Pro 64GB | 48.0 GB | €1,890 / $2,116 | no speed estimate (mixture of experts) | about 7.7 tokens per second, our estimate | no speed estimate (mixture of experts) |
| GMKtec EVO-X2 (Ryzen AI Max+ 395) 64GB | 48.0 GB | €2,199 / $2,035 | no speed estimate (mixture of experts) | about 12.0 tokens per second, our estimate | no speed estimate (mixture of experts) |
| GMKtec EVO-X2 (Ryzen AI Max+ 395) 128GB | 96.0 GB | €3,800 / $3,516 | no speed estimate (mixture of experts) | about 12.0 tokens per second, our estimate | no speed estimate (mixture of experts) |
| Mac Studio M5 Max 128GB | 96.0 GB | €6,000 / $5,551 | no speed estimate (mixture of experts) | about 17.3 tokens per second, our estimate | no speed estimate (mixture of experts) |
Speeds other people measured are self-reported, not ours.
- Qwen3.8-27B on RTX 3090 (18 Sept 2026): RTX 3090, llama.cpp out of the box, about 27.6 tokens per second. community measurement, not ours.
- Qwen3.8-27B on RTX 3090 (18 Sept 2026): RTX 3090, llama.cpp tuned (FA + q8 KV + MTP), about 42.6 tokens per second. community measurement, not ours.
- Qwen3.8-27B on RTX 3090 (18 Sept 2026): RTX 3090, stock vLLM + MTP k=4, about 85.6 tokens per second. community measurement, not ours.
- Qwen3.8-27B on RTX 3090 (18 Sept 2026): RTX 3090, patched vLLM + MTP k=4, about 167.8 tokens per second. community measurement, not ours.
- Qwen3.8-27B on RTX 3090 (18 Sept 2026): RTX 3090, patched vLLM + DFlash2 k=7, about 212.1 tokens per second. community measurement, not ours.
- Josh Gold (killercatfish) (24 Oct 2026): 1× Intel Arc Pro B70 32GB, vLLM (cyspiegel/vllm-xpu-b70), Int4 AutoRound decode [self-reported], about 30 tokens per second. community measurement, single post, not run by us.
- Josh Gold (killercatfish) (24 Oct 2026): 1× Intel Arc Pro B70 32GB, llama.cpp Vulkan (Mesa 26.2.5), Q4_K_M decode [self-reported], about 15 tokens per second. community measurement, single post, not run by us.
- Josh Gold (killercatfish) (24 Oct 2026): 2× Intel Arc Pro B70 32GB split across B850 PCIe 4.0 x4 chipset slot, decode [self-reported], about 5.7 tokens per second. community measurement, single post, not run by us.
- Gabe Ortiz, one RTX 5090: ~220 tokens per second with speculative decoding (plain ceiling ~105 tokens per second). Gabe Ortiz's interview. Self-reported, not ours.
- Roshan George, one RTX A6000 48 GB: about 60 tokens per second with MTP, and about 100 to 129 tokens per second on code with DFlash2. Roshan George's interview. Self-reported, not ours.
- Josh Gold, one Intel Arc Pro B70: see the rows above and Josh Gold's interview. Self-reported, not ours.
- Xijing (Thomas) Wang (@xthomaswang) runs Qwen3.8-27B at 5-bit, served by oMLX with MTP speculative decoding, as the daily backend for Claude Code, Hermes, and his own agent. The machine is a MacBook Pro M5 Max with 128 GB unified memory (about 107 GB of that is GPU-usable). A 64K prompt prefill takes about 110 seconds on that 27B model. A 96 W power adapter drains the battery under long GPU runs. Self-reported, 26 October 2026. Thomas Wang's interview.
- @c4software runs Qwen3.8-27B on two 128 GB Ryzen AI Max+ 395 boxes. @c4software's interview. Self-reported, not ours.
- n8n measured 120 tokens per second for one user and 359 tokens per second across eight users on one NVIDIA DGX Spark (NVFP4, vLLM 0.30.0, 13 Oct 2026). That is n8n's measurement, not ours. Read n8n local LLM hardware requirements.
- Recursive-reported (announcement, not a measurement on this site). With Recursive's own inference engine, Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64GB NVIDIA RTX Spark at short contexts. Source: https://recursive.com/updates/introducing-recursive-flint
Which model should you pick?
- Pick Qwen3.8-27B on one 24 GB card when you want dense quality and spare VRAM. UD-Q4_K_M is 17.80 GB, UD-Q5_K_XL is 22.22 GB, and UD-Q6_K is 23.32 GB, all our estimate.
- Pick Qwen3.6-35B-A3B when you want about 3B active parameters per token. UD-Q4_K_M is 23.10 GB at 8,192 tokens. For a longer agent trace on 24 GB, UD-IQ4_XS is 18.70 GB at 8,192 tokens. See the n8n hardware guide.
- Pick Recursive Flint at UD-Q3_K_XL (19.56 GB) on one 24 GB card. Q4_K_M is 24.46 GB and needs more than 23.4 GB usable. Test it first at the home chat.
- On a 32 GB card such as GeForce RTX 5090 32GB or Intel Arc Pro B70 32GB, and on 64 GB to 128 GB Macs and Strix Halo mini PCs, the standard 4-bit files of all three fit. Our estimate.
Open either Qwen model in the calculator: Qwen3.6-35B-A3B or Qwen3.8-27B. The figures on this page are our estimate.
Related guides: Qwen3.8-27B vs Gemma 4 31B, Gemma 4 31B vs Gemma 4 26B A4B, RTX 3090 vs a $200 a month plan, and one model per GPU vs a tensor split.
Sources and method
Checked . the home chat
Speeds and memory are our estimate from the hardware specs, not a measured test. Prices are dated and may change. How we calculate
Questions
- Should I run Qwen3.6-35B-A3B or Qwen3.8-27B on a 24 GB GPU like the RTX 3090 or RTX 4090?
- Both fit a single 24 GB GPU (23.4 GB usable) at 8,192 tokens, by our estimate. Pick Qwen3.8-27B (16.46 GB measured UD-Q4_K_M file + 0.54 GB conversation memory + 0.80 GB working space = 17.80 GB total) when you want all 27.78B parameters active and spare VRAM for UD-Q5_K_XL (22.22 GB) or UD-Q6_K (23.32 GB). Pick Qwen3.6-35B-A3B (22.13 GB measured UD-Q4_K_M file + 0.17 GB conversation memory + 0.80 GB working space = 23.10 GB total, or 18.70 GB at UD-IQ4_XS) when you want about 3B active parameters per token.
- Does Qwen3.6-35B-A3B fit on a 16 GB GPU (RTX 4060 Ti 16GB, RTX 5080, RX 9060 XT 16GB)?
- Not at the files on our list. The smallest Qwen3.6-35B-A3B download is UD-Q3_K_XL at 17.82 GB total (16.85 GB measured file), and a 16 GB card leaves 15.4 GB usable. Qwen3.8-27B does fit at UD-IQ3_XXS (12.27 GB total, 10.93 GB measured file). Our estimate.
- Why does Qwen3.6-35B-A3B use so little conversation memory at 32K, 128K, and 262K context?
- Only 10 of its 40 layers are full attention. The other 30 are linear attention, which add no growing conversation memory in our estimate. At UD-Q4_K_M that is 0.67 GB at 32,768 tokens (23.60 GB total), 2.68 GB at 131,072 tokens (25.61 GB total), and 5.37 GB at 262,144 tokens (28.30 GB total). Qwen3.8-27B, with 16 full-attention layers, is 17.18 GB at 262,144 tokens (34.44 GB total). Our estimate.
- How does Recursive Flint compare to Qwen3.6-35B-A3B and Qwen3.8-27B on 24 GB VRAM?
- Recursive Flint is 39.79B total and 2.5B active. At Q4_K_M it needs 24.46 GB (23.39 GB measured file + 0.27 GB conversation memory + 0.80 GB working space), which is over the 23.4 GB usable on a 24 GB card. UD-Q3_K_XL is 19.56 GB and fits. Qwen3.6-35B-A3B fits at UD-Q4_K_M (23.10 GB). Qwen3.8-27B fits at UD-Q4_K_M (17.80 GB) and up to UD-Q6_K (23.32 GB). Our estimate.
- Why doesn't localLLMrig show an estimated speed in tokens per second for Qwen3.6-35B-A3B or Recursive Flint?
- Our speed estimate reads every weight for each token. That would understate a mixture of experts, so Qwen3.6-35B-A3B (3B active of 35.95B) and Recursive Flint (2.5B active of 39.79B) show no speed estimate. Qwen3.8-27B is dense, so the hardware pages still publish a tokens-per-second estimate for it. That estimate is not a measured run.
- Can I test Qwen3.6-35B-A3B, Qwen3.8-27B, or Recursive Flint in my browser before downloading a large weights file?
- Recursive Flint is on the free browser chat at https://localllmrig.com/#try, with no login and no local download. Qwen3.6-35B-A3B and Qwen3.8-27B are not on that free list, so the chat cannot run them. Use the calculator links on this page to see the memory estimate before you download a measured file of about 16.46 GB to 22.13 GB.
Want it on paper? The GPU Buying Cheat Sheet (EU, Oct 2026) is a short PDF: graphics memory per budget and the cheapest machine per model size.