Qwen 3.8 started in August 2026 as four models. By the end of September it was nine SKUs, five of which share a backbone with another, three of which did not exist when most "Qwen 3.8 lineup" articles were written, and one of which bills you differently depending on whether you sent text, audio, image or video.
This page is the routing table. It names every Qwen 3.8 variant that actually exists, separates the ones you can download from the ones you can only rent, decodes the naming scheme so you can parse any Qwen model name on sight, and ends with one decision rule. Where a variant has a dedicated guide, it is linked from that variant's row.
What models are in the Qwen 3.8 family?
Nine callable SKUs. Three of them have open weights on Hugging Face; the other six are API-only. Because the two groups answer completely different questions, they get two tables.
Open weights — you can download these. Six repositories under the Qwen organisation, because each model also ships an -FP8 sibling. Three distinct models.
| Model | Params (total / active) | Context | In → out | Licence | Alibaba list price /M | What it's for |
|---|---|---|---|---|---|---|
| Qwen3.8-27B | 27B dense, 64 layers | 262,144 native → 1,000,000 with YaRN | text + image + video → text | Apache 2.0 | $0.50 in / $3.00 out / $0.10 cached | The only member one machine can hold. The default pick. |
| Qwen3.8-Flash-Next | 125B + 51B n-gram + 4B MTP ≈ 180B, 6B active, 512 experts | 262,144 native → 1,000,000 with YaRN | text + image + video → text | Qwen Community License 1.0 | Served as qwen3.8-flash: $0.15 / $0.47 / $0.016 | Throughput economics. Multi-GPU to self-host. |
| Qwen3.8-2.4T-A95B | 2.4T total, 95B active, 92 layers, 512 experts | 262,144 native → 1,010,000 | text → text only | Qwen3.8-Max License | $2 in / $6 out / $0.25 cached | Open-weight flagship. No vision. |
API-only — no weights, ever. These exist as model ids on Qwen Cloud and nowhere else. Nothing in this group has a Hugging Face repository under the Qwen organisation.
| SKU | Backbone | Context | In → out | Price /M (in / out / cached in) | What it's for |
|---|---|---|---|---|---|
| Qwen3.8-Max | 2.4T-A95B | 1M (991K in, 131K out) | text + image + video → text | $2 / $6 / $0.25 | Rolling flagship alias. Adds vision over the open weights. |
| Qwen3.8-Max-0902 | 2.4T-A95B, upgraded snapshot | 1M (991K / 131K) | text + image + video → text | $2 / $6 / $0.25 | Pinned build, 2 Sep 2026. Better long-horizon coding. |
| Qwen3.8-Max-Prime | Max, higher-throughput serving tier | 1M | text + image + video → text | $4 / $12 / $0.50 | Buying latency, not intelligence. See below. |
| Qwen3.8-Flash | Flash-Next plus production features | 1M (991K / 131K, 262K reasoning) | text + image + video → text | $0.15 / $0.47 / $0.016 | The API workhorse. 1M context on by default. |
| Qwen3.8-Omni-Flash | Not published | 1M (991K / 131K, 262K reasoning) | text + image + audio + video → text | $0.15 / $0.47 / $0.016 | Audio and video understanding. Text out only. |
| Qwen3.8-Omni-Flash-Realtime | Not published | Not published (2M TPM, 60 RPM) | text + image + audio + video → text + audio | Per-modality — see the pricing section | Full-duplex live voice. The only SKU that speaks. |
Verified release order: Qwen3.8-27B weights 5 August 2026, Qwen3.8-2.4T-A95B weights 8 August, Qwen3.8-Flash-Next weights 24 August, Qwen3.8-Flash on OpenRouter 26 August, Qwen3.8-Max-0902 2 September (its alias is literally qwen3.8-max-2026-09-02), Qwen3.8-Omni-Flash 21 September, Qwen3.8-Max-Prime 23 September. Qwen3.8-Max itself shipped first, in early August. Omni-Flash-Realtime has no published launch date.
One housekeeping note that trips people up: on OpenRouter, qwen/qwen3.8-max now resolves to qwen3.8-max-0902. The old base listing is gone, so if you pinned the slug you are already on the September snapshot.
Which Qwen 3.8 models can you actually download?
Three: Qwen/Qwen3.8-27B, Qwen/Qwen3.8-Flash-Next and Qwen/Qwen3.8-2.4T-A95B, each with an FP8 twin. That is the complete list. There is no Qwen/Qwen3.8-Max, no Qwen/Qwen3.8-Max-Prime and no Qwen/Qwen3.8-Omni-Flash — searching Hugging Face for those names returns community fine-tunes and distillations with "Max" or "Omni" in the title, not Alibaba releases.
The split has a consequence most lineup tables miss: open weights are the only thing that creates a price market. Count the providers serving each model on OpenRouter today:
| Model | Licence | Providers | Input price spread /M |
|---|---|---|---|
| Qwen3.8-27B | Apache 2.0 | 17, plus a free tier | $0.024 → $0.99. Alibaba's own route sits at $0.425, mid-pack. |
| Qwen3.8-2.4T-A95B | Qwen3.8-Max License | 7 | $2.00 at every single one. Zero price competition. |
| Qwen3.8-Flash-Next / Flash | Qwen Community 1.0 | 1 (Alibaba) | $0.15 — Alibaba's own rate, to the cent. |
| Max, Max-0902, Max-Prime, Omni-Flash | Closed | 1 (Alibaba) | Whatever Alibaba says. |
Apache 2.0 on the 27B produced a twentyfold input-price spread and a free tier. The two custom licences produced seven hosts who all charge exactly the same number, and one host respectively. That is not a coincidence, and the next section explains why.
Is Qwen 3.8 Apache 2.0?
Only the 27B. The family carries three different licences, and they do not get stricter as the models get bigger — they get looser. We read the raw LICENSE files, not the Hugging Face badges, because two of the three show up as the uninformative license: other.
| Model | Licence | Separate licence needed for a MaaS / "AI Work Assistant" business? | Attribution trigger |
|---|---|---|---|
| Qwen3.8-27B | Apache 2.0 | No. Unrestricted commercial use. | None |
| Qwen3.8-Flash-Next | Qwen Community License 1.0 | Yes — at any revenue. No floor at all. | >100M MAU or >US$20M monthly revenue: show the model name in your UI |
| Qwen3.8-2.4T-A95B | Qwen3.8-Max License | Yes, but only above US$50M aggregate revenue over any consecutive 12 months | Same as above |
| Max · Max-0902 · Max-Prime · Flash · Omni-Flash · Omni-Flash-Realtime | Closed, API-only | n/a — Alibaba Cloud terms, no weights to license | n/a |
Read that middle row twice. The smaller, more deployable, easier-to-self-host model is the strictest one in the family. If you run an inference API, a hosted endpoint, a fine-tuning service, or "an independent AI-powered product primarily designed for AI-assisted coding or office productivity" — the licence's own definition, which names Qoder and QwenWork as the archetypes — Flash-Next requires a separate licence from Qwen before any commercial use, with no revenue threshold to sit under. The 2.4-trillion-parameter flagship gives you $50M of headroom before the same clause bites.
Both custom licences carve out internal use: the clause "shall not apply to the licensee's internal Use of the Software, provided that such Use does not make the Software, its outputs, or its underlying model capabilities available to any third party." So an internal coding assistant for your own engineers is fine on either. Reselling inference is where they diverge, and that is exactly the behaviour the provider-count table above measures.
Practical reading: if you are building a product on downloaded weights and you might ever expose the model to customers, the 27B is the only Qwen 3.8 you can ship without a lawyer. Contact for the other two is model-business@notice.qwencloud.com, named in both LICENSE files.
Why do searches for "Qwen 3.8 35B-A3B", "14B" and "122B" come up empty?
Because none of those models exist. This is worth stating flatly, because these names get searched constantly and several of them are real Qwen models — from earlier generations.
- Qwen3.6-35B-A3B is real: 35B total, 3B activated, Apache 2.0, April 2026. It is Qwen 3.6, not 3.8.
- Qwen3.5-122B-A10B is real: 122B total, 10B activated, Apache 2.0, February 2026. It is Qwen 3.5, not 3.8.
- Qwen3.5-35B-A3B and Qwen3.5-9B are also real, also earlier generations.
- There is no Qwen 3.8 at 0.8B, 3B, 9B, 14B, 35B, 70B or 122B. The ladder is 27B dense → ~180B MoE → 2.4T MoE, and nothing in between.
No Qwen/Qwen3.8-35B-A3B and no Qwen/Qwen3.8-122B exist on Hugging Face — a filtered query against the Qwen organisation for either name returns an empty list. What you will find instead are community repacks like Qwen3.8-2B-Heretic-Max and various Qwen3.8-27B-…-GGUF quantisations, none of them Alibaba releases.
What happened is that Qwen changed its size ladder between generations. 3.5 and 3.6 offered mid-size MoE models a workstation could hold; 3.8 skipped that band entirely. If you were running Qwen3.6-35B-A3B, the 3.8 successor is Qwen3.8-27B — fewer total parameters, but dense rather than 3B-active, so meaningfully stronger per token despite the smaller number. Qwen's own 27B model card frames it as building "upon the 3.6-27B version". If you were running Qwen3.5-122B-A10B there is no direct successor at that size; our Qwen 3.5 complete guide still covers it properly.
What do the Qwen 3.8 name parts mean — A3B, MoE, Flash, Max, Next, Omni, Prime, Realtime?
Qwen names are compositional. Once you know the pieces you can parse any of them on sight, including the ones that do not exist yet.
| Token | Meaning |
|---|---|
3.8 | Generation. Qwen ships point releases roughly quarterly — 3.5 in February 2026, 3.6 in April, 3.7, 3.8 in August. |
27B, 122B, 2.4T | Total parameters. Determines how much memory the weights occupy. |
A3B, A10B, A95B | "Activated" — parameters that actually run per token. Appears only on MoE models. Sets speed and price, not memory. No Qwen 3.8 model uses an A-suffix name, which is why "Qwen 3.8 A3B" returns nothing. |
| MoE | Mixture of Experts. Both Flash-Next and 2.4T-A95B hold 512 experts and fire 10 routed plus 1 shared per token. You hold all 512 in memory; you 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. |
Next | Architecture-preview line, and in 3.8 it also marks the open-weight build. Flash-Next carries the newer pieces — n-gram embeddings, Qwen Sparse Attention, Gated Residual, a multi-token-prediction head. Qwen's own card says qwen3.8-flash "is the official version based on Qwen3.8-Flash-Next with more production features, e.g., 1M context length by default, official built-in tools." |
Max | The flagship. Hosted, most capable, most expensive. |
Omni | Input modality breadth, not output. Omni adds audio to the text/image/video inputs the rest of the family already takes. It does not mean the model talks. |
Prime | Throughput, not intelligence. A higher-capacity serving tier for the same model at a higher price. This is the single least guessable suffix in the family — see the next section. |
Realtime | Full-duplex streaming endpoint with audio output and per-modality billing. A different API shape, not just a faster one. |
0902 / date suffixes | A pinned snapshot. qwen3.8-max-0902 = alias qwen3.8-max-2026-09-02. Undated ids are rolling and can change under you. |
FP8 | An 8-bit release of the same weights. Roughly halves serving memory versus BF16. Not a different model. |
Instruct / Thinking | Retired in 3.8. Every Qwen 3.8 model is hybrid: thinking is on by default, depth set with reasoning_effort (low, medium, xhigh — xhigh is the default). An "Instruct" or "Thinking" suffix on something calling itself Qwen 3.8 is a community repack. |
One exception to the reasoning_effort story: on Qwen3.8-2.4T-A95B thinking cannot be switched off. Its model card is explicit — "a text-only model that requires thinking mode for all interactions. Multimodal inputs are not supported, and thinking cannot be disabled." Every response opens with a reasoning block, and reasoning tokens bill as output.
What is Qwen3.8-Max-Prime and is it worth double the price?
Max-Prime is a serving tier, not new weights. OpenRouter's description, verbatim: "a higher-throughput variant of Qwen3.8 Max from Alibaba's Qwen team, served as a separate SKU at a higher price point."
Six independent checks say the same thing:
- It does not appear in Qwen Cloud's model catalogue at all, while Max-0902 and Flash both do.
qwencloud.com/models/qwen3.8-max-primereturns 404.- There is no Hugging Face repository for it under the Qwen organisation.
- Artificial Analysis has no page for it —
artificialanalysis.ai/models/qwen3-8-max-primeis a 404 — while Max, 27B and 2.4T-A95B all have one. - The Qwen3.8-Max model page itself never mentions a Prime tier.
- The pricing doubles uniformly. $2 → $4 input, $6 → $12 output, $0.25 → $0.50 cached read. Exactly 2.00× on all three axes. Capability upgrades do not price like that; capacity reservations do.
Context that makes it make sense: Artificial Analysis measures Qwen3.8 Max (0902) at 37.1 output tokens per second, ranking #180 of 224 models on output speed, with 2.64 s to first token (AA measurement, October 2026). Max is genuinely slow. A throughput SKU is a targeted fix for a measured weakness, not a marketing exercise — and it means the only reason to pay for Prime is that your workload is latency-bound. If it is throughput-bound and batchable, Max-0902 at half the price gives you the same answers.
Full workup on both new SKUs: Qwen3.8 Max-Prime and Omni-Flash explained.
Does Qwen3.8-Omni-Flash support audio output?
No. Omni-Flash takes text, image, audio and video in, and returns text only. Qwen Cloud's own documentation is blunt about it: "Qwen3.8-Omni-Flash supports text output only. Do not set audio."
What "Omni" buys you is input breadth. It is the first Qwen model built around agentic use with native audio-video understanding, and it handles two-channel and four-channel spatial audio as input — useful for meeting recordings, multi-mic captures and video with positional sound. Its context limits match Flash exactly (1M window, 991K max input, 131K max output, 262K max reasoning) and so does its price, to the cent: $0.15 in, $0.47 out, $0.016 cached.
If you need the model to speak, you need the separate Qwen3.8-Omni-Flash-Realtime endpoint, which outputs text and audio, supports full-duplex audio-video interaction, understands audio in over 60 languages and speaks in over 30. It is the only SKU in the family that generates audio.
What does each Qwen 3.8 model cost?
Alibaba's published list prices, per million tokens:
| SKU | Input | Output | Cached input |
|---|---|---|---|
| Qwen3.8-Flash | $0.15 | $0.47 | $0.016 |
| Qwen3.8-Omni-Flash | $0.15 | $0.47 | $0.016 |
| Qwen3.8-27B | $0.50 | $3.00 | $0.10 |
| Qwen3.8-Max · Max-0902 · 2.4T-A95B | $2.00 | $6.00 | $0.25 |
| Qwen3.8-Max-Prime | $4.00 | $12.00 | $0.50 |
Omni-Flash-Realtime is the outlier: it bills per modality, so your cost depends on what you send, not just how much.
| Omni-Flash-Realtime — billed component | Price /M tokens |
|---|---|
| Audio input | $0.93 |
| Text input | $0.70 |
| Text / image / video input | $0.23 |
| Text & audio output | $1.87 — and output text alone is not charged |
Two things to notice. Audio input at $0.93 costs four times the image/video rate on the same endpoint, so a voice agent's bill is driven by how much the user talks, not by video frames. And because text-only output is free on Realtime, a pipeline that uses Realtime for listening and emits text costs less than you would guess from the headline $1.87.
One honest caveat on cross-referencing prices: OpenRouter's Alibaba-routed rates run about 15% under Alibaba's own list. It shows the 27B at $0.425 / $2.55 against Alibaba's published $0.50 / $3.00 — the same 0.85 multiplier on input, output and cache. Treat vendor list price as the ceiling and check your actual route.
Which Qwen 3.8 model should you use?
The rule in one sentence: if it has to run on your hardware, use 27B; if it runs over an API, use Flash unless you have measured that it fails your task, in which case use Max-0902; and only reach for Prime, Omni-Flash or Realtime when you need latency, audio input or speech out specifically.
| Situation | Use | Why |
|---|---|---|
| Local, single 24 GB GPU (4090, 3090, 5090) | Qwen3.8-27B, Q4_K_M | 18 GB on Ollama, leaves room for context. The only 3.8 that fits. |
| Local, 16 GB GPU | Qwen3.8-27B at Q3 | Tight but workable with reduced context. |
| Apple Silicon workstation | qwen3.8:27b-mlx | Ollama ships an MLX tag at the same 18 GB. Comfortable from 32 GB unified memory up. |
| Shipping a commercial product on downloaded weights | Qwen3.8-27B, full stop | Apache 2.0. The other two need a separate Qwen licence if you expose the model to third parties. |
| Cheap API at volume | Qwen3.8-Flash | $0.15 / $0.47 with 1M context. Nothing else in the family is close. |
| Hardest reasoning and long agentic runs | Qwen3.8-Max-0902 | Highest measured intelligence in the family, pinned so it won't shift under you. |
| Flagship quality, weights required | Qwen3.8-2.4T-A95B | Same backbone family, text-only, but you own it — and the $50M licence floor is generous. |
| Latency-bound flagship workload | Qwen3.8-Max-Prime | Only reason to pay 2×. If you can batch, don't. |
| Meeting notes, podcast analysis, video with sound | Qwen3.8-Omni-Flash | Audio in at Flash prices, including spatial audio. Text out. |
| Live voice agent | Qwen3.8-Omni-Flash-Realtime | The only SKU that emits audio. Budget for the $0.93 audio-input rate. |
Install paths and per-GPU tuning for the local options: how to run Qwen 3.8 locally.
How do the Qwen 3.8 models compare on benchmarks?
Vendor-reported figures from Alibaba's own model cards, so directional rather than independent. Two benchmarks are named identically across all three open-weight cards, which makes them the only fair columns.
| Model | SWE-bench Pro | GPQA Diamond | Other vendor figures |
|---|---|---|---|
| Qwen3.8-27B | 61.7 | 89.2 | Terminal Bench 2.1 73.0 · OSWorld-Verified 84.3 · WebArena-Verified 64.8 · AndroidWorld 81.9 · LiveCodeBench v6 90.3 |
| Qwen3.8-Flash-Next | 62.5 | 91.7 | SWE-bench Multilingual 81.0 · AndroidWorld 84.5 · LiveCodeBench v6 91.9 · HLE 35.9 |
| Qwen3.8-Max (2.4T-A95B) | 67.7 | 92.6 | Terminal Bench 2.1 86.6 · PaperBench 93.0 · FrontierSWE 73.5 · IFBench 82.8 |
Independent third-party numbers add a detail the model cards cannot. On Artificial Analysis' Intelligence Index v4.3.2 (October 2026), Qwen3.8 Max (0902) scores 45.42, the open-weight Qwen3.8-2.4T-A95B scores 39.89, and Qwen3.8-27B scores 33.70 at its default xhigh reasoning effort. Those first two share a backbone, so the 5.5-point gap is the 0902 snapshot pulling ahead of the released weights. "Max equals 2.4T-A95B" was true in August and is now only approximately true — the hosted route has moved on, and the open weights have not been re-released to match. Quote those three numbers with the version attached: AA rebased the Index from v4.1.1 to v4.3.2 in October 2026, and because the eval basket changed rather than the models, v4.1.1-era scores do not convert onto this scale.
Output speed from the same source: 27B 45.2 tok/s at xhigh, 2.4T-A95B 37.7, Max (0902) 37.1. The shape of the family is that intelligence and speed trade off almost perfectly against each other, which is the gap Max-Prime is sold into.
Dedicated write-ups with the full benchmark tables and harness caveats: Qwen3.8-27B, Qwen3.8 Flash, Qwen3.8 Max.
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, Qwen 3.6 had 35B-A3B. Qwen 3.8 has neither. The ladder is 27B dense, then ~180B, then 2.4T.
Instruct and Thinking merged. One checkpoint per size with a reasoning_effort dial, which halves the number of models you have to evaluate.
Modality became a product axis instead of a repo suffix. Earlier generations shipped separate -VL repositories. In 3.8, vision is standard on everything except the open-weight 2.4T-A95B, and the extra modalities — audio in, audio out — are sold as distinct SKUs rather than distinct checkpoints. That is why the SKU count grew from four to nine without a single new set of weights.
Version-by-version detail, including whether upgrading from 3.6 is worth the migration: Qwen 3.8 vs Qwen 3.6.
FAQ
How many Qwen 3.8 models are there?
Nine callable SKUs as of October 2026: Qwen3.8-27B, Qwen3.8-Flash-Next, Qwen3.8-Flash, Qwen3.8-2.4T-A95B, Qwen3.8-Max, Qwen3.8-Max-0902, Qwen3.8-Max-Prime, Qwen3.8-Omni-Flash and Qwen3.8-Omni-Flash-Realtime. Only three of those are distinct sets of weights, and the Hugging Face repository count is six because each open-weight model also ships an FP8 twin.
Which Qwen 3.8 models have open weights?
Exactly three: Qwen3.8-27B, Qwen3.8-Flash-Next and Qwen3.8-2.4T-A95B, each with an FP8 sibling repository under the Qwen organisation on Hugging Face. Everything with Max, Omni or Realtime in the name is API-only, and so is the production qwen3.8-flash endpoint — though that one is built on the downloadable Flash-Next weights.
Is Qwen 3.8 Apache 2.0?
Only Qwen3.8-27B. Qwen3.8-Flash-Next ships under the Qwen Community License 1.0 and Qwen3.8-2.4T-A95B under its own Qwen3.8-Max License. Counter-intuitively, Flash-Next is the stricter of the two: it requires a separate licence from Qwen for any Model-as-a-Service or AI-coding-assistant business with no revenue threshold, while the 2.4T flagship only triggers the same clause above US$50M of revenue over 12 months. Both carve out purely internal use.
What is Qwen3.8-Max-Prime?
A higher-throughput serving tier for Qwen3.8 Max, not a more capable model. It prices at exactly double Max on input, output and cached reads ($4 / $12 / $0.50), has no Hugging Face repository, does not appear in Qwen Cloud's catalogue, and has no Artificial Analysis page. Buy it only if your workload is latency-bound — Artificial Analysis clocks Max at 37.1 output tokens per second, #180 of 224 as of October 2026, so the slowness it fixes is real.
Does Omni-Flash support audio output?
No. Qwen3.8-Omni-Flash accepts text, image, audio and video but returns text only — Qwen Cloud's documentation says "supports text output only. Do not set audio." For speech output you need Qwen3.8-Omni-Flash-Realtime, which emits text and audio, handles full-duplex audio-video interaction, and bills per modality rather than at a flat rate.
Which Qwen 3.8 should I use?
Qwen3.8-27B if it must run on your own hardware or ship inside a commercial product (it is the only Apache 2.0 member). Qwen3.8-Flash for API work at volume, at $0.15 per million input tokens. Qwen3.8-Max-0902 for the hardest reasoning. Omni-Flash only if you need audio input, Realtime only if you need audio output, and Max-Prime only if you are latency-bound.
Is there a Qwen 3.8 35B-A3B, 14B or 122B?
No. Qwen3.6-35B-A3B (April 2026) and Qwen3.5-122B-A10B (February 2026) are real models from earlier generations, and there has never been a Qwen 3.8 at 0.8B, 3B, 9B, 14B, 35B, 70B or 122B. The generation has nothing between 27B and ~180B. Upgrading from Qwen3.6-35B-A3B, the intended successor is Qwen3.8-27B — smaller on paper, but dense rather than 3B-active.
What is the smallest Qwen 3.8 model?
Qwen3.8-27B, at 27 billion dense parameters — about 18 GB at Q4_K_M, which is what ollama pull qwen3.8 gives you. There is no 4B, 8B or 14B in this generation. If you need something smaller you have to drop back to a Qwen 3.6 or 3.5 checkpoint, or quantise the 27B harder.
So which one should you actually pull?
Start with Qwen3.8-27B. It is the only Apache 2.0 member, it fits on hardware you probably already own, it takes images and video, and its gap to the flagship is about six points of SWE-bench Pro. Run it for a week on your real workload before spending anything.
When you outgrow your own machine, move to qwen3.8-flash rather than Max — thirteen times cheaper on input for roughly a five-point coding difference. Pin qwen3.8-max-0902 rather than the rolling qwen3.8-max alias for the specific tasks where you have measured Flash failing. Treat Max-Prime, Omni-Flash and Omni-Flash-Realtime as what they are: three answers to three narrow questions — latency, audio in, audio out. None of them is an upgrade to the model you are already using.
If you are putting these models into production and want engineers who have already shipped LLM infrastructure, Codersera can extend your team with vetted remote developers.