Qwen 3.8 Model Lineup: Every Variant and Which to Use
Qwen 3.8 is a small family with confusing names. Alibaba shipped it across August 2026 in four public forms, and the parameter counts in those names range from 27 billion to 2.4 trillion — a factor of nearly 90. Getting the choice wrong means either buying hardware you did not need or trying to run a model that will never fit on your machine.
This page is the routing table. It names every Qwen 3.8 variant that actually exists, decodes the naming scheme so you can parse any Qwen model name on sight, and gives you one decision rule. Where we have a dedicated guide for a variant, it is linked from that variant's section.
What models are in the Qwen 3.8 family?
Four public models, three of which have open weights. Every open-weight model also ships an -FP8 sibling repository, so the Hugging Face count is six repositories under the Qwen organisation, not six distinct models.
| Model | Params (total / active) | Architecture | Context | Open weights | Licence | Price or hardware floor | Best for |
|---|---|---|---|---|---|---|---|
| Qwen3.8-27B | 27B / 27B | Dense, vision encoder | 262,144 native, ~1M with YaRN | Yes | Apache 2.0 | 18 GB at Q4_K_M → one 24 GB GPU. API from ~$0.43 / $2.55 per M | Local use; the default pick |
| Qwen3.8-Flash-Next | 180B / 6B active | Sparse MoE, vision encoder | 262,144 native, ~1M with YaRN | Yes | qwen-community-1.0 | $0.15 / $0.47 per M. Self-hosting needs a multi-GPU server | Cheap API at volume |
| Qwen3.8-2.4T-A95B | 2.4T / 95B active | Sparse MoE, text only | 262,144 native, ~1,010,000 extended | Yes | Custom qwen3.8-max licence | $2 / $6 per M. Self-hosting is multi-node | Maximum open-weight capability |
| Qwen3.8-Max | Same backbone as 2.4T-A95B | Hosted, accepts text + image + video | 1,000,000 | No | Alibaba Cloud terms | $2 / $6 per M | Flagship with vision, managed |
Release order on OpenRouter, which is the cleanest public timeline: Max on 3 August 2026, 2.4T-A95B on 12 August, 27B on 14 August, Flash-Next on 26 August. All four expose a 1M-token context on the hosted route.
Two details in that table trip people up. First, Qwen3.8-Max and Qwen3.8-2.4T-A95B are the same model with different packaging — the open-weight release even carries a licence literally named qwen3.8-max. But they are not interchangeable: the open-weight card states plainly that "multimodal inputs are not supported", while the hosted Max endpoint accepts text, image and video. If you download the weights you get the text backbone only.
Second, Flash-Next's 180B is not 180B of transformer. The model card breaks it down as 125B in the main network with 6B activated per token, plus 51B of n-gram embedding and 4B of multi-token-prediction head. Six billion active parameters is why it prices at a tenth of Max.
Why do searches for "Qwen 3.8 35B-A3B" and "Qwen 3.8 122B" come up empty?
Because neither model exists. This is worth stating flatly, because both names get searched constantly and both are real Qwen models — from earlier generations.
- Qwen3.6-35B-A3B is real. 35B total, 3B activated, 256 experts with 8 routed plus 1 shared, Apache 2.0, released April 2026. It is Qwen 3.6, not 3.8.
- Qwen3.5-122B-A10B is real. 122B total, 10B activated, 256 experts, Apache 2.0, released February 2026. It is Qwen 3.5, not 3.8.
There is no Qwen/Qwen3.8-35B-A3B and no Qwen/Qwen3.8-122B on Hugging Face. Searching for the former surfaces only a community fine-tune (Lord-H4D3ZS/Qwen3.8-Distill-35B-A3B-Coder-Abliterated), which is a 3.6-shaped model distilled from 3.8 outputs — not an Alibaba release.
What happened is that Qwen changed its size ladder between generations. Qwen 3.5 and 3.6 offered mid-size MoE models in the 35B–122B range that hobbyists could run on a workstation. Qwen 3.8 skipped that band entirely: it goes 27B dense, then straight to 180B, then to 2.4T. If you were running Qwen3.6-35B-A3B and want the 3.8 equivalent, the answer is Qwen3.8-27B — fewer total parameters, but dense rather than 3B-active, so it is meaningfully stronger per token despite the smaller number in the name.
If you were running Qwen3.5-122B-A10B, there is no direct 3.8 successor at that size. Flash-Next is the closest in spirit — sparse, cheap per token — but it is a served model, not a workstation one. Our Qwen 3.5 complete guide still covers the 122B-A10B properly if that is the model you actually meant.
What does the naming mean — A3B, MoE, Flash, Next, Max, FP8?
Qwen names are compositional. Once you know the pieces you can parse any of them.
| Token | Meaning |
|---|---|
3.8 | Generation. Higher is newer. Qwen ships point releases roughly quarterly — 3.5 in February 2026, 3.6 in April, 3.8 in August. |
27B, 122B, 2.4T | Total parameter count. This is what determines how much memory the weights occupy. |
A3B, A10B, A95B | "Activated". The number of parameters that actually run per token. A3B = 3 billion active. It appears only on mixture-of-experts models and it is what determines speed and price, not memory. |
| MoE | Mixture of Experts. The network is split into many expert sub-networks and a router picks a handful per token. Qwen3.8-2.4T-A95B has 512 experts and fires 10 routed plus 1 shared. You must hold all 512 in memory but you only pay compute for 11. |
| Dense | No routing — every parameter runs on every token. Qwen3.8-27B is dense, which is why it has no A suffix. |
Flash | The speed-and-price tier. Fewest active parameters, cheapest per token, first choice for high-volume work. |
Next | Architecture-preview line. Flash-Next carries the newer pieces — n-gram embeddings, a multi-token-prediction head — that have not yet propagated to the rest of the family. |
Max | The flagship. Hosted, most capable, most expensive. |
FP8 | An 8-bit release of the same weights. Roughly halves serving memory versus BF16 with minor quality cost. Not a different model. |
Instruct / Thinking | Older Qwen split: chat-tuned versus reasoning-tuned. Qwen 3.8 has retired this split. All Qwen 3.8 models are hybrid — thinking is on by default and controlled with a reasoning_effort parameter (low, medium, xhigh). If you see an Instruct or Thinking suffix on something calling itself Qwen 3.8, it is a community repack. |
One exception to the reasoning_effort rule: on Qwen3.8-2.4T-A95B thinking cannot be turned off at all. Every response begins with a reasoning block. That matters for latency-sensitive products and for anyone metering output tokens, since reasoning tokens bill as output.
Which Qwen 3.8 model should you use?
The decision rule, in one sentence: if it has to run on your own hardware, use 27B; if it runs over an API, use Flash-Next unless you have measured that it fails your task, in which case use Max.
That covers most cases because the 27B is the only member of the family a single machine can hold, and Flash-Next is roughly thirteen times cheaper than Max on input tokens while scoring close to it on published reasoning benchmarks.
| Situation | Use | Why |
|---|---|---|
| Local, single 24 GB GPU (RTX 4090, 3090, 5090) | Qwen3.8-27B at Q4_K_M | 18 GB download, leaves room for context. The only 3.8 that fits. |
| Local, 16 GB GPU | Qwen3.8-27B at Q3 | ~13 GB with reduced context. Tight but workable. |
| Apple Silicon workstation | Qwen3.8-27B, MLX tag | Ollama ships qwen3.8:27b-mlx. Unified memory handles the 18 GB comfortably from 32 GB up. |
| Local, multi-GPU server | Qwen3.8-Flash-Next-FP8 | 6B active makes it fast once resident, but you need to hold ~180B of weights. |
| Cheap API at scale | Qwen3.8-Flash-Next | $0.15 / $0.47 per M with 1M context. Nothing else in the family is close. |
| Maximum capability, vision needed | Qwen3.8-Max | Only route that takes image and video at flagship quality. |
| Maximum capability, weights required | Qwen3.8-2.4T-A95B | Same backbone, text only, but you own the weights. |
| Coding agent, long sessions | Flash-Next first, 27B locally | Both are vision-capable and both score in the low 60s on SWE-bench Pro. |
What is Qwen3.8-27B good at?
It is the family's centre of gravity and the only one most people will ever run themselves. 27 billion dense parameters across 64 layers, using a hybrid attention stack that alternates three Gated DeltaNet blocks against one gated-attention block. Native context is 262,144 tokens, extensible toward 1M with YaRN scaling. Apache 2.0, so commercial use is unrestricted.
Vendor-reported scores put it at 61.7 on SWE-bench Pro, 89.2 on GPQA Diamond, 84.3 on OSWorld-Verified for computer use and 64.8 on WebArena-Verified. It accepts text, image and video input.
Practically: 18 GB at the default Q4_K_M quantisation, 30 GB at Q8_0, 56 GB at BF16. Ollama carries only this member of the family — ollama pull qwen3.8 gets you the 27B and nothing else, which is worth knowing before you go looking for a Flash tag that does not exist.
Full write-up: Qwen3.8-27B complete guide. For the install path and per-GPU tuning, see how to run Qwen 3.8 locally.
What is Qwen3.8-Flash-Next good at?
Throughput economics. 180B of weights, 6B firing per token, 512 experts with 10 routed plus 1 shared, 48 layers, native 262K context extending to 1M. Licence is qwen-community-1.0 rather than Apache 2.0, so read it before shipping a commercial product on the downloaded weights.
Its published numbers are not a step down from the 27B — 62.5 on SWE-bench Pro and 91.7 on GPQA Diamond, both marginally ahead — while costing $0.15 per million input tokens. It also handles text, image and video, and scores 84.5 on AndroidWorld for on-device agent tasks.
The catch is self-hosting. Six billion active parameters is cheap to compute but you still have to keep all 180B resident, which puts it firmly in multi-GPU-server territory. For nearly everyone this is an API model. It appeared on OpenRouter on 26 August 2026. Dedicated guide: Qwen3.8 Flash complete guide.
What are Qwen3.8-Max and 2.4T-A95B good at?
Hard reasoning and long agentic runs. 2.4 trillion total parameters, 95 billion active, 92 layers, 512 experts. Native 262K context extending to about 1,010,000 tokens. Vendor-reported: 67.7 on SWE-bench Pro, 92.6 on GPQA Diamond, 93.0 on PaperBench, 86.6 on Terminal Bench 2.1 — though our Qwen 3.8 Max guide notes that neutral-harness re-runs of that Terminal-Bench figure land closer to 81.
Choose between the two forms on two axes. Vision: only the hosted Max takes images and video. Ownership: only 2.4T-A95B gives you weights, under a custom licence that is permissive but is not Apache 2.0 — check the revenue clauses if you are a large company. Price is identical either way at $2 / $6 per million tokens, and Max additionally offers cached-read pricing at $0.25 per million.
How do the three open-weight models compare on benchmarks?
These are Alibaba's own published figures from the model cards, so treat them as directional rather than independent. Two benchmarks are named identically across all three cards, which makes them the only fair columns to compare.
| Model | SWE-bench Pro | GPQA Diamond | Input $/M |
|---|---|---|---|
| Qwen3.8-27B | 61.7 | 89.2 | ~0.43 |
| Qwen3.8-Flash-Next | 62.5 | 91.7 | 0.15 |
| Qwen3.8-2.4T-A95B | 67.7 | 92.6 | 2.00 |
The shape of that table is the argument for Flash-Next as the default API model: it costs a thirteenth of the flagship and gives up about five points of SWE-bench Pro. The flagship earns its price on the hardest agentic work and on anything where a five-point coding gap compounds across a long run.
What changed from Qwen 3.6 and Qwen 3.5?
Three structural changes, all visible in the naming.
The mid-size MoE band disappeared. Qwen 3.5 had 122B-A10B and Qwen 3.6 had 35B-A3B. Qwen 3.8 has neither. The ladder is now 27B dense, then 180B, then 2.4T.
Instruct and Thinking merged. Earlier generations shipped separate reasoning-tuned checkpoints. Qwen 3.8 has one checkpoint per size with a reasoning_effort dial, which halves the number of models you have to evaluate.
Vision became standard. Qwen3.8-27B, Flash-Next and Max all accept image and video input without a separate VL repository. Only the open-weight 2.4T-A95B is text-only.
Version-by-version detail, including whether an upgrade from 3.6 is worth the migration: Qwen 3.8 vs Qwen 3.6.
FAQ
How many Qwen 3.8 models are there?
Four public models: Qwen3.8-27B, Qwen3.8-Flash-Next, Qwen3.8-2.4T-A95B and the hosted Qwen3.8-Max. Three of those have open weights on Hugging Face, and each open-weight model also has an FP8 repository — which is why you will see six Qwen3.8 repositories under the Qwen organisation rather than three.
What does A3B mean in Qwen model names?
"A" stands for activated. A3B means roughly 3 billion parameters run per token, out of a much larger total. It appears only on mixture-of-experts models: Qwen3.6-35B-A3B holds 35B of weights but fires about 3B per token. Active parameters set speed and price; total parameters set memory. No Qwen 3.8 model uses the A3B suffix.
Which Qwen 3.8 model is best?
Qwen3.8-Max (equivalently 2.4T-A95B) is the most capable, leading the family on SWE-bench Pro at 67.7 and GPQA Diamond at 92.6. But "best" depends on the constraint: Qwen3.8-27B is the best model you can run on one consumer GPU, and Flash-Next is the best value on an API at $0.15 per million input tokens.
Is Qwen 3.8 open source?
Partly. Qwen3.8-27B is Apache 2.0, which is genuinely permissive. Qwen3.8-Flash-Next ships under a custom qwen-community-1.0 licence and Qwen3.8-2.4T-A95B under a custom licence named qwen3.8-max — both open-weight but with terms you should read before commercial deployment. Qwen3.8-Max itself is API-only with no weights released.
What is the smallest Qwen 3.8 model?
Qwen3.8-27B, at 27 billion dense parameters. There is no 4B, 8B or 14B in this generation, unlike earlier Qwen releases. If you need something smaller than 27B, you have to drop back to a Qwen 3.6 or Qwen 3.5 checkpoint, or run the 27B at a more aggressive quantisation — Q3 brings it to about 13 GB.
Qwen 3.8 Flash vs 27B — which should I use?
Use 27B if the model must run on your own hardware; it is the only Qwen 3.8 that fits on a single GPU. Use Flash-Next if you are calling an API, because it is cheaper per token and scores slightly higher on both SWE-bench Pro and GPQA Diamond. Flash-Next needs a multi-GPU server to self-host, so it is not a local alternative.
Is there a Qwen 3.8 35B-A3B or a Qwen 3.8 122B?
No. Qwen3.6-35B-A3B (April 2026) and Qwen3.5-122B-A10B (February 2026) are real models from earlier generations. Qwen 3.8 has no model between 27B and 180B. If you are upgrading from Qwen3.6-35B-A3B, the intended successor is Qwen3.8-27B — smaller on paper, but dense rather than 3B-active.
So which one should you actually pull?
Start with Qwen3.8-27B. It is Apache 2.0, it fits on hardware you probably already have, it takes images, and its benchmark gap to the flagship is about six points of SWE-bench Pro. Run it for a week on your real workload before spending anything.
If you then need more throughput than your own machine gives you, move to Flash-Next on an API rather than to Max — the price difference is more than tenfold and the capability difference is small. Reserve Max, or the 2.4T-A95B weights, for the specific tasks where you have measured Flash-Next failing. That sequence costs the least and tells you the most.
If you are building on these models in production and want engineers who have already shipped LLM infrastructure, Codersera can extend your team with vetted remote developers.