TL;DR: Qwen3.8-Flash-Next is a 125B-A6B MoE released August 26, 2026 under the Qwen Community License 1.0 — not Apache 2.0, unlike Qwen3.8-27B. A Q4-class GGUF is 108–111GB, so it does not fit any consumer GPU. If you have 128GB of unified memory it runs fast; everyone else should stay on Qwen3.8-27B or rent.

	The short answer
License	Qwen Community License 1.0 — not Apache 2.0 (checked Oct 2026)
Architecture	MoE: 125B total / 6B active, plus 51B n-gram embedding + 4B MTP head
Context	262,144 tokens native, extendable to 1M
Weights	~180GB-class BF16; Q4-class GGUF 108–111GB
Hardware floor	~75GB RAM/unified memory minimum (Unsloth); 128GB to be comfortable
Ollama	Library tags exist (125b-a6b-q4_K_M, q8_0, bf16) as of late Oct 2026
Fits a 24GB GPU?	No. Not at any usable quant.

Honest take: Unless you already own a 128GB unified-memory box, skip this model. Qwen3.8-27B is Apache 2.0, fits a used RTX 3090, and remains the sane default for one person self-hosting.

What license does Qwen3.8-Flash-Next use?

Qwen3.8-Flash-Next ships under the Qwen Community License 1.0, not Apache 2.0. Hugging Face lists the license as “other,” and the official README punts with “please find the license file released with the model weights on Hugging Face Hub or ModelScope.” That alone tells you it isn’t a standard OSI license.

The terms that matter, from the license text as republished in the Ollama library (read the LICENSE file on the model repo yourself before any commercial deployment — huggingface.co was unreachable from our environment this run, so we could not read the canonical file directly):

• Scale clause: commercial products with more than 100 million monthly active users or US$20 million in monthly revenue must prominently display the model’s name in the UI.
• MaaS carve-out: running a “Model as a Service” or “AI Work Assistant” business requires a separate license from Qwen. Internal use is exempt as long as outputs and model capabilities are not offered to third parties.
• Third-party summaries also list attribution and no-competing-model-training conditions; we could not confirm those against the primary text, so verify before relying on them.

For a home lab or internal company tool, none of this bites. For anyone reselling inference, it does. This breaks Qwen’s own pattern — smaller Qwen3-family checkpoints like Qwen3.8-27B are plain Apache 2.0, and the Reddit threads that announced Flash-Next assumed the same. They were wrong, the same way the Apache 2.0 assumption was wrong for Qwen3.8-Max.

How big is Qwen3.8-Flash-Next, really?

The official model card says 125B total parameters with 6B activated per token, plus a 51B n-gram embedding table and a 4B multi-token-prediction (MTP) head. That’s why trackers like Artificial Analysis report “180B”: the full on-disk footprint is the main model plus the embedding table and draft head.

Verified sizes as of October 2026:

Variant	Size	Source
BF16 full weights	~360GB	llmrun.dev calculator
Unsloth UD-Q4_K_XL GGUF	111.33GB	Unsloth quant table, checked Oct 9 2026
Community Q4_K_M GGUF	~108–111GB (3 shards)	community HF repos
Unsloth stated minimum	≥75GB RAM or unified memory	Unsloth docs

The 75GB minimum assumes an aggressive sub-Q4 dynamic quant. For Q4-class quality you need the weights (~111GB) plus KV cache plus OS overhead — in practice a 128GB machine, loaded carefully.

One architectural detail helps multi-GPU setups: the README states the 51B n-gram embedding table “can be offloaded to host memory” with asynchronous prefetching. That cuts the GPU-resident portion to roughly the 125B main model (~70GB at Q4), which is still four 24GB cards, not one.

Does Qwen3.8-Flash-Next fit a 24GB GPU?

No. A Q4-class quant is 108–111GB — about 4.5× the VRAM of an RTX 3090 or 4090. There is a third-party claim that a 4.27bpw build runs on a 10GB card by spilling everything into system RAM; technically llama.cpp will do that, but you’d be streaming ~100GB of weights through a PCIe bus and system RAM you probably don’t have. It’s a demo, not a setup.

Realistic tiers:

• 128GB unified memory (Strix Halo boxes, Mac Studio): the practical single-box floor for Q4-class quants.
• 96GB unified memory or 4×24GB GPUs with the embedding table offloaded to host RAM: workable with Unsloth’s dynamic quants.
• 24GB or less: run Qwen3.8-27B instead. It’s Apache 2.0 and its Q4_K_M is ~17GB.

How fast is it once it fits?

Fast — capacity is the problem, not speed. Decode speed is bound by memory bandwidth divided by bytes read per token, and an MoE only reads its active parameters: 6B active at Q4 is roughly 3.5–3.7GB per token. On a Strix Halo’s 256GB/s that’s a theoretical ceiling around 70 tok/s (expect 70–95% of ceiling in practice, so ~50–65 tok/s); on a Mac Studio M5 Ultra’s 1.2TB/s bandwidth the model is effectively instant for chat use. Compare that with a dense 70B at Q4, which reads 42.5GB per token and crawls at 4–6 tok/s on the same Strix Halo. These are calculated ceilings, not our benchmarks — treat vendor and community tok/s claims above the ceiling as wrong.

How do you run it?

Ollama now carries the model — library tags including 125b-a6b-q4_K_M, 125b-a6b-q8_0, and 125b-a6b-bf16 are live as of late October 2026 (an Oct 9 hardware guide found no tag yet, so this landed mid-October; ollama.com was unreachable from our environment this run, so confirm the exact tag before a 100GB download):

ollama pull qwen3.8-flash-next:125b-a6b-q4_K_M
ollama run qwen3.8-flash-next:125b-a6b-q4_K_M
# Expect a ~105-111GB download. Check `ollama ps` for actual memory use.

llama.cpp supports the architecture (text and vision) per the official README; grab a GGUF from Hugging Face and point llama-server at the first shard.

vLLM / SGLang are the official multi-GPU paths. The repo’s own commands assume tensor parallelism across 4 GPUs:

vllm serve Qwen/Qwen3.8-Flash-Next --port 8000 \
  --tensor-parallel-size 4 --max-model-len 262144 \
  --reasoning-parser qwen3 --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

Both expose an OpenAI-compatible endpoint, so Open WebUI and Ollama-style local stacks front it without changes.

What to actually buy

Prices as of October 2026, verified against current street prices — not MSRPs:

Your situation	The move	Price	Where
24GB card, want Apache 2.0	Used RTX 3090 + Qwen3.8-27B	~$1,456–$1,558	Check price
Want Flash-Next on one quiet box	GMKtec EVO-X2 128GB (Strix Halo)	~$3,499–$3,649	Check price
Want headroom for bigger MoEs + 1M context	Mac Studio M5 Ultra (up to 512GB)	from $5,499	Check price
Undecided — test the workload first	Rented multi-GPU 4090 box	from ~$0.14/hr per card	Vast.ai

Hardware math for the unified-memory tier lives on our sister site: Ryzen AI Max+ 395 / Strix Halo guide.

When NOT to self-host Qwen3.8-Flash-Next

• You’re building a paid inference or AI-assistant product. The MaaS carve-out means you need a separate license from Qwen. Apache 2.0 alternatives (Qwen3.8-27B, GLM-5.3-Flash’s MIT 320B-A18B) don’t carry that call-a-lawyer step.
• You have less than 96GB of memory. Sub-Q4 quants exist, but quantization degrades knowledge non-linearly below Q4 — you’re paying 100GB of disk for degraded answers a 27B at Q4 may match.
• Your workload is bursty. At ~$3,500 for the cheapest comfortable box, occasional use is cheaper rented: a 4-card 4090 box on Vast.ai at ~$0.56/hr covers ~6,200 hours before matching the hardware price — years of evenings-and-weekends use.
• You need audited, reproducible terms. “Other” licenses with revenue triggers change between model generations; Qwen has now used three different license types across the 3.8 family alone.

FAQ

Is Qwen3.8-Flash-Next open source? It’s open-weight, not open source. The Qwen Community License 1.0 permits local and internal commercial use but gates MaaS businesses and adds display obligations past 100M MAU / $20M monthly revenue. Apache 2.0 within the same family means Qwen3.8-27B.

Can I run Qwen3.8-Flash-Next on an RTX 4090? Not on one. Q4-class weights are 108–111GB. Four 24GB cards with the 51B embedding table offloaded to host RAM is the minimum discrete-GPU setup; one 128GB unified-memory machine is simpler.

Is Qwen3.8-Flash-Next on Ollama? Yes — qwen3.8-flash-next library tags (125b-a6b-q4_K_M, q8_0, bf16) appeared in the Ollama library between October 9 and late October 2026. Verify the tag on ollama.com/library before starting the ~105GB+ pull.

Sources

• Qwen/Qwen3.8-Flash-Next — Hugging Face model card (125B/6B + 51B n-gram + 4B MTP; license “other”)
• QwenLM/Qwen3.8-Flash-Next — official GitHub (release 2026-08-26, serve commands, embedding offload)
• Unsloth docs: Qwen3.8-Flash-Next (≥75GB minimum, UD-Q4_K_XL 111.33GB)
• llmhardware.io hardware requirements guide (checked Oct 9 2026)
• Ollama library: qwen3.8-flash-next (quant tags, license text)

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe