Qwen3.8-Flash-Next memory tiers: what fits in 16GB / 64·96GB / planned 128GB

Bottom Line Qwen3.8-Flash-Next, released 26 August 2026, is not the Qwen3.8-27B in the previous note. Official card: 125B backbone (6B active) + 51B n-gram embeddings + 4B MTP. Compute is sparse. Disk and resident size are workstation-class. Unsloth GGUF is about 72.5–111GB; community MLX 4-bit is 103.8GB. On today’s fleet (4080 16GB / M1 Max 64GB / M2 Max 96GB) the question is not “does it all fit in VRAM,” it is where n-gram and expert weights live. AMD AI Max+ 395 128GB and Gigabyte DGX Spark 128GB are still not here. This article has no tok/s from us.

This is not a copied bench table. It is where the files go, and where they stop, when AI Tech Engine tries Flash-Next — shipped the same week — on the fleet we actually have.

Official card — how it differs from 27B

Repo: Hugging Face Qwen/Qwen3.8-Flash-Next. MarkTechPost and community summaries list the license as qwen-community-1.0 (not Apache 2.0). Check the HF LICENSE yourself before commercial use or redistribution.

	Official text	How we read it
Parameters	125B (6B active) + 51B n-gram + 4B MTP	“6B active” is FLOPs. Disk/resident is an ~180B-class package
Type	Causal LM + Vision Encoder	Time text path and vision-on separately
Context	Native 262,144, YaRN up to 1,000,000	Full 262K is not an assumption
Layout	48 layers, `12 × (3 × (GDN→MoE) → 1 × (QSA→MoE))`	Sparse attention, MoE, and n-gram together
N-gram	20,000,000 bigram/trigram entries (layer 2)	Assumes host RAM/SSD offload. Not an accelerator-only problem
Serving	vLLM, SGLang, TokenSpeed, transformers, llama.cpp (GGUF)	Daily loop is llama.cpp / (Mac) MLX. We do not copy multi-GPU FP8 recipes as our result

Checkpoint sizes cited by MarkTechPost: FP8 172.78 GiB, BF16 335.28 GiB. Sparse activation only cuts compute. It does not shrink storage.

We do not copy Qwen’s own bench tables or the “1/9 training cost” claim. The comparison set is the Qwen3.8-27B mapping and the M1 Max 27B timing. Flash-Next speed cells are still empty.

Community quant file sizes (not our measurements)

We did not re-measure these. They are document / card figures we use when planning a load.

Unsloth GGUF (`unsloth/Qwen3.8-Flash-Next-GGUF`, run guide)

quant	listed size
UD-IQ1_S	72.5 GB
UD-IQ1_M	74.5 GB
UD-Q2_K_XL	78.9 GB
UD-IQ3_XXS	82.0 GB
UD-Q3_K_XL	90.0 GB
UD-IQ4_XS	93.7 GB
UD-Q4_K_XL	111.3 GB

Unsloth’s docs write “75GB RAM / unified memory, GPU VRAM not required” together with n-gram (PLE) SSD/host offload. That is their guide. It is not a promise the same numbers appear on our boxes.

llama.cpp side: community reports that qwen4exp / Flash-Next paths merged around 2026-08-27. This lab has not recorded a build commit or flags yet.

Community MLX (PipeNetwork table)

build	listed size
bfloat16 (upstream)	360.0 GB
8-bit	192.2 GB
6-bit	148.0 GB
mixed-4/8-bit	106.2 GB
4-bit	103.8 GB

The repo calls mixed-4/8-bit the build to use under 150GB, and says uniform 4-bit takes a large PPL hit (+20.6% vs bf16). Recommended MLX builds do not fit whole on M1 Max 64GB / M2 Max 96GB. Lines like ~354GB resident bf16 and ~24 tok/s are not our hardware. We do not copy them.

Split GGUF with n-gram on SSD (reference)

AtomicChat and others published builds that park n-gram shards on SSD and lower the in-memory ask (example: ~84.9–110.5GB total, ~45.8–56.1GB in-memory as listed). Even when a card says “64GB/96GB Mac guide,” this lab has not verified load, quality, or speed. We do not mix “someone claims it works” with “our needle.”

Fleet mapping — today vs planned

Numbers mean “is there somewhere for the file to live?”, not speed.

Today (on site)

Path	RTX 4080 16GB	M1 Max 64GB MacBook	M2 Max 96GB Mac Studio
Official FP8/BF16 weights	No (172–335 GiB class)	No	No
Unsloth GGUF fully in RAM	No. 16GB VRAM ≪ 72–111GB	Tight to no even at 1-bit (72GB+). Their 75GB floor is already under 64GB	Borderline. Low-bit GGUF + short context is a candidate. Q4-class 111GB overflows
GGUF + n-gram/expert offload (llama.cpp CUDA/Metal)	Experiment candidate. Card holds active experts + KV; rest and n-gram on system RAM/SSD. To write “it ran,” quant, offload flags, and pass/fail travel together	Experiment candidate. Metal + mmap/SSD exists in the community; unmeasured here	Experiment candidate. More headroom than 64GB. Also unmeasured
MLX 4-bit / mixed (~104–106GB)	Not a target	Not recommended. File exceeds 64GB	Not recommended. File overflows 96GB or leaves no KV/vision room
Vision on	Tighter. Text first	Same	Same

The 4080 16GB point is not “Flash-Next in VRAM.” It is whether 16GB compute + large system RAM can even smoke n-gram offload. That cell is still blank.

Planned (not arrived)

Path	AMD AI Max+ 395 128GB Strix Halo (planned)	Gigabyte DGX Spark 128GB (planned)
Unsloth Q4-class GGUF (~93–111GB)	Target tier for weights + leftover KV	Same. CUDA/llama.cpp, vLLM if it applies
Runbook with n-gram on NVMe	ROCm/Vulkan path after arrival	Community Spark runbooks already exist. No install diary before it lands
MLX	Not Apple	Not Apple
Buy / availability (late-August 2026 snapshot)	Strix Halo 128GB mini-PCs mix sold-out and expensive reports on DRAM supply. We do not treat a price list as a bench	Founders/reseller MSRP bands are public; this lab has not confirmed an order or arrival

We do not quote planned-box nameplates (200B inference and so on) as Flash-Next results.

How this relates to the 27B note

	Qwen3.8-27B mapping	This note (Flash-Next)
Architecture	Dense 27B (+vision)	MoE 125B-A6B + 51B n-gram
4-bit file scale	Unsloth ~14–18GB	Unsloth ~94–111GB (near Q4)
Workhorse today	The box it loads on (M1/M2 Max)	On boxes we hold, offload / split is the default assumption
Our tok/s	M1 Max timing exists	None

Flash-Next is not “27B mapped again.” A same-week release opens the fleet’s empty cells (128GB) and offload experiments.

What this lab will measure next (Flash-Next)

Before 128GB arrives, only on hardware we hold.

Load profile — same Unsloth GGUF on 4080 (CUDA offload) / M1 Max / M2 Max. Success, OOM, resident memory, flags used.
Where the n-gram sits — RAM vs SSD (mmap). Do the listed in-memory savings reproduce on our boxes.
Quality smoke — not just load. Short coding and instruction-following without collapse/loops. (No pasted bench scores.)
Vision off, first — text. mmproj and video after that.
After 128GB arrives — same files on Strix Halo and Spark; Q4-class resident + long context on the same needle.

What we will not do: treat Qwen/Unsloth/MLX repo tok/s as ours, extrapolate M5 Max / 5090 community speed onto 4080 / M1 / M2, write planned 128GB as already measured, or reuse the 27B timing table as a Flash-Next result.

File sizes and tier mapping are checked as of today. The speed column stays empty.