Qwen3.6-35B-A3B VRAM requirements

Yes, tight — and the build choice decides by how much. The ggml-org Q4_K_M (the official llama.cpp-org build) totals 20.1 GiB at 32K context: 19.02 GiB weights + ≈0.5 GiB runtime overhead + 0.625 GiB config-derived KV, leaving 2.4 GiB of a 24GB card's 22.5 GiB usable — under this site's 2.5 GiB tight threshold. The unsloth UD-Q4_K_M (MTP build) weighs 21.11 GiB and totals 22.2 — only 0.3 GiB to spare. The densest fit is unsloth's UD-Q4_K_XL at 22.4 GiB (0.1 to spare). NVIDIA's own NVFP4 build does not fit at 32K: 22.9 GiB all-in misses usable by 0.4 GiB (at 8K context it totals 22.47 GiB — a knife-edge that fits by 0.03). Above Q4 nothing fits 24GB at 32K: UD-Q5_K_S already totals 24.9 GiB, 2.4 GiB over usable.

16GB is the smallest sane tier: at 32K context an RTX 4060 Ti 16GB tops out at UD-IQ3_XXS (14.2 GiB all-in, 0.3 GiB to spare) — no 4-bit build fits 16GB (UD-IQ4_XS totals 18.1). At 8K context the picture barely changes: the KV term shrinks to ≈0.16 GiB, and UD-IQ3_XXS drops to 13.7. A 12GB card has no ceiling on the builds this page enumerates: the smallest of them (UD-IQ1_M, the 1-bit MTP build at 10.59 GiB of weights) totals 11.7 GiB at 32K and 11.2 GiB at 8K — both over the 10.5 GiB usable on an RTX 3060 12GB. The cheapest tracked file overall is the non-MTP unsloth UD-IQ1_M at 9.36 GiB (pinned, in the builds table sources): at 8K context that file totals ≈10.0 GiB and fits a 12GB card, tight; at 32K it is a 0.02-GiB knife-edge — arithmetically inside, practically pinned to the metal, and this page does not recommend it as a real 12GB door. For a real 12GB route see the Strata engine question below (reported hybrid mode).

No — activation is not residency. Every token touches only 3.0B of the 35.9B parameters (8 of 256 experts, plus 1 shared), which is what makes the MoE fast; but all 35.9B parameters must sit in VRAM for any of it to work. Weights alone are 19.0 GiB at Q4_K_M — a 36B-class memory bill with a 3B-class compute bill. VRAM follows total parameters, always. (This is the same rule the predecessor Qwen3-30B-A3B page states for its 30.5B/3.3B split.)

20.1 GiB at 32K context: 19.02 GiB weights + ≈0.5 GiB runtime overhead + 0.625 GiB KV cache. The KV figure is config-derived, not a community measurement: 20,480 bytes/token × 32K, counting only the 10 full-attention GQA layers (2 KV heads × 256 head dim, bf16) — the 30 GatedDeltaNet layers hold a constant recurrent state instead of per-token KV, so the hybrid bill grows ~4.8× slower per context token than the all-attention predecessor. Context scaling at Q4_K_M: ≈19.6 GiB at 4K, 19.7 at 8K, 22.0 at 128K (still inside a 24GB card, tight), and ≈24.5 at the full 262,144 — past usable, so the full native window needs a 32GB card. Strata's doc states the same shape in fp32 ('≈45 KB/token' for the 10 attention layers plus the MTP block).

Same quant name, two different files. unsloth ships two repos: Qwen3.6-35B-A3B-MTP-GGUF (with the multi-token-prediction draft layer compiled in, for speculative decoding) and Qwen3.6-35B-A3B-GGUF (without). The MTP layer adds ≈0.4-1.6 GiB per file: UD-Q4_K_XL is 21.28 GiB with MTP vs 20.82 without, Q8_0 is 35.21 vs 34.37, UD-Q4_K_M is 21.11 vs 20.61. The Strata engine's dedicated qwen35 path pulls the MTP builds (reported, linked below); plain llama.cpp/Ollama/LM Studio runs don't need the draft layer — take the smaller non-MTP file unless your runtime does speculative decoding. ggml-org's repo instead ships the MTP weights as separate mtp-* sidecar files, which this site excludes from weight sums (same rule as mmproj vision packs).

Per Strata's own documentation — reported, we link it, we did not re-run it: yes, in hybrid mode. Strata's dedicated engine path (src/qwen35) loads the unsloth MTP builds (UD-Q4_K_XL / UD-Q6_K_XL / Q8_0) and, on an RTX 3060 12GB, keeps 22-37 of the 256 experts per layer in VRAM with the rest on CPU — a tiered door this site's plain-fit ladder doesn't price. On an R9700 32GB every expert stays in VRAM (decode 117.8 tok/s on 0.1.45, reported), and parallel serving runs up to 8 slots (R9700 8 requests: 325.4 tok/s aggregate vs llama.cpp's 208.1, reported). The honest half of the claim: 'The Mac engine does not run this model.' If you're on Apple Silicon, the plain GGUF builds via llama.cpp-class runtimes are the route — see the 48GB-class door note above. Engine home: Strata.

Both are 24GB-tier Q4_K_M cards (the related row below prices the 27B; this page's hero is 20.1 GiB at 32K). The difference is how they spend the budget: the 27B is the dense sibling — every parameter works on every token; the 35B-A3B is a 256-expert MoE that touches only 3.0B parameters per token, which is where its decode speed comes from, at the cost of carrying 35.9B parameters resident. Same card runs either; the MoE buys tok/s, the dense model buys simplicity (no expert-routing behavior to reason about). This is a same-family comparison — the 27B page carries its own measured ladder.

The successor changes the architecture, not the card tier. The predecessor is 30.5B parameters (3.3B active) with 48 full-attention GQA layers — its KV bill is 98,304 B/token, a reported 3.0 GiB at 32K. This model is 35.9B (3.0B active) with only 10 full-attention layers among its 40 (30 GatedDeltaNet linear layers handle the rest), so its config-derived KV is 20,480 B/token = 0.625 GiB at 32K — the long-context bill grows ~4.8× slower. Both land in the same place for a buyer: Q4_K_M-class GGUF, ~20-21 GiB all-in, fits a 24GB card tight. Both natively span 262,144 tokens. The predecessor's page answers the same questions for the older architecture — linked in the related row below (and this page is linked back from it).

Weights are measured file sizes at pinned revisions (sources above); totals add runtime overhead (0.4951 GiB, the llama.cpp anchor used site-wide) plus the KV cache at 32K context. Here KV@32K = 0.625 GiB, config-derived: 20,480 B/token on the 10 full-attention layers (2 KV heads × 256 dim × 2 bytes × K,V), provisional until a community measurement lands — the derivation inputs live in the db and the method box under the quantization table. A build fits a card at 32K when the total clears usable VRAM (nominal minus reserve), and tight means under 2.5 GiB of headroom. The 30 GatedDeltaNet layers hold a constant recurrent state (~0.03 GiB config-derived) instead of per-token KV — disclosed, below page rounding, not counted.