Qwen3.8-Flash-Next VRAM requirements

Alibaba's hybrid-MoE release of August 26, 2026 (public release date per same-week coverage; the HF repo was created August 24, and the model card points at an official technical report). The pinned config: 48 layers in repeating groups of three Gated-DeltaNet linear layers and one full-attention QSA layer, 512 experts per layer with 10 routed plus 1 shared active, a 256K context, an MTP layer, vision input, and the 20M-entry n-gram embedding table at layer 2. The model card's own parameter line reads 'Number of Parameters: 125B with 6B activated, plus 51B n-gram embedding and 4B MTP' — in this site's caliber that is a ~125.0 B language body with ~6.0 B active, plus the 51.2B-parameter n-gram table, and the config's model_type — qwen4_exp — is the release-name mapping for the Qwen4-Exp architecture, which is how external tooling refers to it. License: Qwen Community License 1.0 (commercial use with attribution; two carve-outs quoted in the deployment section and pinned below). Supply at the October 20, 2026 sweep: 52 measured builds across eight publishing orgs — the official 131-shard BF16 release; GGUF ladders from unsloth, ISTA-DASLab, bartowski, ggml-org, and lmstudio-community; NVFP4 datacenter builds from nvidia and RadixArk.

You will see three numbers for its size, and all three are correct at their own boundary. The language body — attention stacks, experts, shared expert, embeddings — is ~125B parameters: the model card's headline, and our config arithmetic lands within a rounding of it. Add the n-gram embedding table — 20M entries × 2560 dims = 51.2B parameters, bigger than most standalone open models — and the total is ~175B. Count the MTP layer and the vision tower as well and the widest caliber is ~180B, which is what the HF page tends to auto-display. The table is not an appendix: it participates at layer 2 of every forward pass, and it is inside every file sum priced on this page. So, one caliber sentence for comparing apples to apples: the body is ~125B parameters, the total is ~175B including the 51.2B n-gram embedding table — when a source says '125B' or '180B', check which boundary it counts first.

No — and not on any 24–48GB card. The smallest measured build is ISTA-DASLab's GSQ-RCO Q2_0 at 62.4 GiB total (2× the 4090's 24GB, 2× the 5090's 32GB), and every one of the 52 measured builds sits above the 48GB class. No published build file fits a 24–48GB card — but two engine arrangements run it there anyway, both reported rather than files: Strata tiers the raw weights across RAM, VRAM, and SSD behind a stated 32GB-combined door, and a community SGLang recipe (sgl-project issue #37792, September 3, 2026) serves a 2-bit AutoRound conversion on one 24GB GPU with 32GB of host RAM at ~54–57 tok/s decode (reported, linked below). The working exits: a 128GB appliance runs standard builds Q2_0 through Q4_K_M; one 192GB B200 covers the same builds for renters; a 512GB Mac Studio runs everything measured including official BF16; or the API.

One 192GB B200 covers Q2_0 through Q4_K_M — from 62.4 GiB up to 111.9 GiB — and official BF16 (335.8 GiB) needs 2. Card counts are plain ceiling division on measured totals (pinned below); a full 5× B200 node is a throughput choice, not a capacity requirement. MI300X nodes divide the same way at the same per-GPU memory class. Renting one B200 to serve a 6B-active MoE is usually worse economics than the API unless residency is the point — this is the one case where the appliance section beats the cloud section.

Yes — both boxes, and without lossy conversions. The 128GB class (~120 GiB usable; M4 Max / DGX Spark included) fits standard builds Q2_0 (62.4 GiB) through Q4_K_M (111.9 GiB) with 8-plus GiB of spare at the top — the same GGUF files from the rent table, no special runtime. The 512GB M3 Ultra (480 GiB usable) runs every measured build including the official BF16 release at 335.8 GiB. The caveat that applies to every unified-memory row on this site: bandwidth trails datacenter GPUs by an order of magnitude, so treat these as 'it runs' verdicts — long-context prefill is where you'll feel it. KV stays small next to these weights: config-derived figures put 32K of cache under a GiB even before the linear layers' constant state is counted.

No official distill or smaller sibling is tracked in our sources as of October 20, 2026 — and with this architecture (12 full-attention + 36 linear layers, 512 experts per layer, the n-gram table) a step-down is a new training run, not a quant. Two real answers exist. Model step-down: Qwen3.8-27B, the same family's 27B-body champion, priced for 24GB-class cards. Hardware step-down: the Strata engine (product wrapper, all claims reported) tiers these exact raw weights across RAM, VRAM, and SSD — only 10 of each layer's 512 experts resident at a time — and states a door of 32GB of RAM and VRAM together with a single 12GB+ card. That is the pattern this model started: the community's answer to the cliff was streaming the weights, not shrinking them.

The GGUF supply is broad and current — five publishing orgs as of October 20, 2026, including ggml-org (the llama.cpp project itself) and lmstudio-community (the LM Studio ecosystem channel); that breadth is the loader-support signal this site tracks. nvidia's NVFP4 build is the datacenter-native outlier — a Blackwell-oriented safetensors release for B200-class deployment rather than a desktop file. The official 131-shard BF16 safetensors is the Transformers path. An Ollama library listing is outside this page's enumerated sources as of October 20, 2026 — if Ollama support matters to you, check the library directly. Whichever path you pick, re-verify against the pinned trees below before budgeting hardware.

Provider-dependent — check before you code against it. Our OpenRouter snapshot of October 18, 2026 enumerates the qwen3.8 siblings (qwen3.8-flash, qwen3.8-27b, the qwen3.8-max line) but no flash-next listing, so the family search linked below may or may not resolve to these exact weights at a given provider. This page answers for the files; the API section links the search. Where a provider does serve it under an alias, the provider's live model list — not the name — is the source of truth for which build you get.