Qwen3.8-Flash-Next is Alibaba’s answer to a question the whole industry has been circling for a year: how small can the active part of a frontier model get before the quality falls over? The answer the Qwen team published on 26 August 2026 is six billion parameters — out of a hundred and twenty-five billion sitting on disk.
That ratio is the entire story. Qwen3.8-Flash-Next is a multimodal mixture-of-experts model with 125 billion total parameters, an additional 51 billion parameters held in a novel N-gram embedding table, and just 6 billion parameters activated for any given token. Alibaba open-weighted it on Hugging Face and ModelScope the same evening, alongside an FP8 checkpoint, and described it in the release notes as “an early preview of the architecture used in Qwen4”.
We have covered the rest of this family as it shipped — the enormous Qwen3.8-Max-Preview in July and the Qwen3.8 27B open-weight release in August — and the wider catalogue lives in our AI models hub. This one is different in kind. It is not a bigger model or a smaller model. It is a different machine, shipped early so the community can pull it apart before Qwen4 is built on top of it.
What follows is what Alibaba actually published, what the architecture does, where the benchmark table is strong and where it quietly is not, what the thing costs to rent and to run, and which parts of the announcement have not been independently verified by anybody.
Table of contents
- What Alibaba Released With Qwen3.8-Flash-Next
- Inside the Qwen3.8-Flash Architecture
- The 125B MoE Numbers Behind Qwen3.8-Flash
- What Multimodal Means for Qwen3.8-Flash
- Benchmarks: Qwen3.8-Flash Against Opus 4.6 and DeepSeek-V4
- The One-Ninth Training Cost Claim
- Long Context and Speed in Qwen3.8-Flash
- Qwen3.8-Flash Pricing Against the Rest of the Lineup
- Running Qwen3.8-Flash-Next on Your Own Hardware
- The Licence Attached to Qwen3.8-Flash-Next
- Why Qwen3.8-Flash-Next Previews Qwen4
- What Qwen3.8-Flash Means for Technology Buyers
- Risks and Open Questions
- Frequently Asked Questions
- References and Further Reading
What Alibaba Released With Qwen3.8-Flash-Next
The release was a weights drop with a technical report attached, not a product launch, and the distinction matters for how you read the numbers.
The timing
The Qwen team scheduled the open-sourcing for 11pm Beijing time on 26 August 2026. A ModelScope listing appeared and was pulled the day before, which is how the specification leaked ahead of the announcement — the 125B-A6B configuration was circulating in developer forums roughly twenty-four hours before Alibaba confirmed it.
What is in the repository
Two checkpoints shipped: the BF16 weights and an FP8 quantisation, both under the Qwen/Qwen3.8-Flash-Next namespace, with a matching GitHub repository carrying deployment instructions. Community quantisations followed within hours, including a full GGUF ladder from Unsloth.
The framing Alibaba chose
The Qwen team’s own description is unusually modest for a model that beats Claude Opus 4.6 on most of its published comparisons. Qwen3.8-Flash-Next is presented as a preview — the architecture that Qwen4 will use, shipped ahead of the model family that will exploit it properly. That framing is a hedge, and it is also probably accurate.
Where the production version sits
The open weights carry the -Next suffix. The production hosted model drops it: Qwen3.8-Flash will be served through the QwenCloud API at $0.16 per million input tokens and $0.47 per million output tokens. At the time of the weights release that API had been announced but not opened, so the only way to run the model on day one was to run it yourself.
Release-day facts
| Item | Detail |
|---|---|
| Model | Qwen3.8-Flash-Next (open weights); Qwen3.8-Flash (hosted) |
| Published | 26 August 2026, 11pm Beijing time |
| Total parameters | 125 billion, plus a 51 billion N-gram embedding table |
| Active per token | 6 billion |
| Context window | 262,144 tokens native, 1,000,000 with YaRN |
| Modalities | Text, image and video in; text out |
| Checkpoints | BF16 and FP8, plus community GGUF and NVFP4 |
| Licence | qwen-community-1.0 |
| Serving stacks | vLLM, SGLang, TokenSpeed, Transformers |
Inside the Qwen3.8-Flash Architecture
Four subsystems were rebuilt, and each one attacks the same problem from a different angle: how to add capacity without adding compute.
Hybrid attention: Gated DeltaNet plus QSA
The previous generation paired Gated DeltaNet with conventional gated attention. Qwen3.8-Flash-Next replaces the second half with Qwen Sparse Attention. Gated DeltaNet compresses history cheaply through a linear recurrence; QSA then uses a lightweight compressed indexer to decide which parts of that history are worth reading at full precision.
Why QSA works on blocks, not tokens
The important design choice in QSA is granularity. Rather than scoring every individual token for retrieval, it selects at the level of micro-blocks, with a published budget of 512 blocks or 2,048 tokens. Block-level selection is what makes the indexer cheap enough to run on a 262,144-token context without eating the savings the sparsity was supposed to deliver.
The layer layout
The model card spells the stack out: twelve repetitions of three Gated DeltaNet-plus-MoE layers followed by one QSA-plus-MoE layer, for 48 layers total. So only a quarter of the layers pay for real attention. The remainder run linear-time recurrence, which is why the long-context arithmetic works at all.
Gated Residual
The residual stream is widened into four parallel branches, with a bottleneck rank of 320. An element-wise, data-dependent read gate decides what each branch sees, and a per-branch scalar write gate decides what it contributes back. In plain terms: more paths for information to travel, with learned traffic control on each one, which is a training-stability play as much as a capability one.
N-gram embedding, the genuinely new part
This is the component with no obvious precedent at scale. A 51-billion-parameter table, 20 million entries wide, holds embeddings keyed on local bigrams and trigrams, injected at layer 2. It adds enormous capacity for almost no floating-point cost, because a table lookup is not a matrix multiply.
The part that makes it practical
Crucially, that table does not have to live on the GPU. It can be offloaded to host memory and prefetched asynchronously, overlapped with computation, so the 51 billion parameters cost you system RAM rather than accelerator memory. That single engineering decision is what keeps a 176-billion-parameter object servable on four GPUs.
The optimiser and the training recipe
Muon and AdamW are applied to different weight categories rather than one optimiser being used throughout, with Muon refined for orthogonalisation accuracy and parameter splitting. Guided by refitted scaling laws, the team also removed batch-size warmup entirely and started training at the target batch size, cutting total optimiser steps substantially.
The specification at a glance
| Component | Configuration | What it buys |
|---|---|---|
| Layers | 48, as 12 x (3 x GDN-MoE + 1 x QSA-MoE) | Only 25% of layers use full attention |
| Hidden dimension | 2,560 | Narrow trunk, capacity moved to experts |
| Experts | 512 total, 10 routed + 1 shared active | About 2% of experts fire per token |
| Expert intermediate size | 640 | Many small experts, not few large ones |
| Gated DeltaNet heads | 48 value, 16 query-key, head dim 128 | Cheap linear history compression |
| QSA heads | 24 query, 2 key-value, head dim 256 | Small KV cache per layer |
| QSA budget | 512 blocks / 2,048 tokens | Fixed attention cost at any context length |
| Gated Residual | 4 branches, bottleneck rank 320 | Training stability at high sparsity |
| N-gram embedding | 51B parameters, 20M entries, layer 2 | Capacity without matrix multiplication |
| Token embedding | 248,320 padded vocabulary | Broad multilingual coverage |
The 125B MoE Numbers Behind Qwen3.8-Flash
Sparsity ratios have become the headline specification of 2026, and this one is the most aggressive yet published by a major lab.
Two per cent of the experts, five per cent of the weights
Ten routed experts plus one shared expert fire from a pool of 512. That is roughly two per cent of the expert population per token. Counting the whole model, 6 billion active parameters against 125 billion total works out at 4.8 per cent — so more than nineteen twentieths of the model sits idle on any given forward pass.
The comparison that frames it
Qwen3.7-Plus, the model this one is measured against, carries 397 billion total parameters and activates 17 billion per token. Qwen3.8-Flash-Next activates 6 billion. That is 35 per cent of the active compute, against a total parameter count that is 31 per cent as large. Both axes shrank by roughly two thirds at once.
The 4B multi-token prediction head
A separate 4-billion-parameter multi-token prediction module ships with the model. MTP heads are what make speculative decoding work without a draft model, and their presence is a strong hint that the serving story — not just the training story — was designed in from the start.
Where the memory actually goes
The honest accounting is that this is a 176-billion-parameter artefact: 125 billion in the backbone plus 51 billion in the N-gram table. The reason nobody describes it that way is that the table is designed to sit in host RAM. It is a real cost, just not a real GPU cost.
Active versus total, four models compared
What Multimodal Means for Qwen3.8-Flash
The word gets used loosely. Here it means the model accepts text, images and video as input, and the agentic scores suggest that capability is doing real work rather than sitting on the specification sheet.
The inputs it accepts
Text, image and video all go in; text comes out. There is no image generation and no audio path. For most enterprise use the video input is the interesting one, because screen recordings and captured workflows become directly promptable rather than needing a separate extraction pipeline.
AndroidWorld is the number to look at
On AndroidWorld, a benchmark that drives a real Android environment through a sequence of interface actions, Qwen3.8-Flash-Next scores 84.5 against 62.0 for Claude Opus 4.6. A 22.5-point gap on a task that requires reading a screen and acting on it is the single largest margin anywhere in the published table.
Multimodal reasoning
ClawEval-MM returns 64.4 at Pass@3, against 60.4 for the dense Qwen3.8-27B release. That is a four-point improvement from a model activating a fraction of the compute, and it is the cleanest like-for-like evidence that the sparse architecture has not cost anything on the vision side.
The caveat worth stating
Multimodal benchmarks are less mature than coding benchmarks, the harnesses vary more between labs, and AndroidWorld in particular is sensitive to how the environment is configured. A 22-point margin is impressive and should still be reproduced locally before anyone builds a product around it.
Benchmarks: Qwen3.8-Flash Against Opus 4.6 and DeepSeek-V4
The published table puts Qwen3.8-Flash-Next ahead of Claude Opus 4.6 on the majority of the benchmarks the two were compared on. That deserves both the attention it is getting and rather more scepticism than it is getting.
Software engineering
On SWE-bench Pro the model records 62.5 against 53.4 for Claude Opus 4.6 and 54.4 for DeepSeek-V4-Flash. On SWE-bench Multilingual it reaches 81.0 against 77.5 for Opus 4.6. On DeepSWE 1.1 it posts 58.7 against 54.4 for DeepSeek-V4-Flash. Three coding benchmarks, three wins, none of them narrow.
Agentic and office work
CoWorkBench returns 73.9 against 68.2 for Opus 4.6 and 45.1 for DeepSeek-V4-Flash. JobBench returns 55.7 against 36.6 for Opus 4.6 and 27.6 for Qwen3.7-Plus — roughly double its own predecessor. Toolathlon Verified sits at 73.5. This cluster, not the coding scores, is where the generational jump is largest.
Knowledge and reasoning
GPQA Diamond comes in at 91.7 against 91.3 for Opus 4.6 and 90.8 for DeepSeek-V4-Flash — a three-way tie inside half a point, which is what a saturated benchmark looks like. LiveCodeBench v6 reaches 91.9 against 88.8. IFBench, measuring instruction following, records 81.3 against 62.5 for Opus 4.6.
Where it loses
Humanity’s Last Exam is the exception, and Alibaba published it anyway: 35.9 for Qwen3.8-Flash-Next against 40.0 for Claude Opus 4.6. HLE rewards deep, broad, hard-to-compress knowledge, which is precisely what a model with a 2,560 hidden dimension and 6 billion active parameters would be expected to struggle with. The architecture’s weakness shows up exactly where theory says it should.
How to read the whole table
These are vendor-published numbers, run on the vendor’s harness, released the same day as the weights. Nothing about them is disqualifying — the weights are public, so anyone can check — but nothing about them is independent either. The correct posture is the one any serious AI strategy applies to a supplier’s own benchmark: treat it as a hypothesis, not a finding.
The published comparison
| Benchmark | Qwen3.8-Flash-Next | Claude Opus 4.6 | Margin |
|---|---|---|---|
| SWE-bench Pro | 62.5 | 53.4 | +9.1 |
| SWE-bench Multilingual | 81.0 | 77.5 | +3.5 |
| CoWorkBench | 73.9 | 68.2 | +5.7 |
| JobBench | 55.7 | 36.6 | +19.1 |
| AndroidWorld | 84.5 | 62.0 | +22.5 |
| IFBench | 81.3 | 62.5 | +18.8 |
| LiveCodeBench v6 | 91.9 | 88.8 | +3.1 |
| GPQA Diamond | 91.7 | 91.3 | +0.4 |
| Humanity’s Last Exam | 35.9 | 40.0 | -4.1 |
The four largest margins
The One-Ninth Training Cost Claim
The efficiency assertion attached to this release is the one that will matter longest, and it is also the one nobody outside Alibaba can check.
What the claim says
The Qwen team states that Qwen3.8-Flash-Next achieves better results than Qwen3.7-Plus at roughly one-ninth the training cost, with the largest gains in coding and office tasks. Both halves of that sentence are load-bearing, and only one of them is verifiable from the weights.
Why it is plausible
A 397-billion-parameter model activating 17 billion per token, trained conventionally, against a 125-billion model activating 6 billion with batch-size warmup removed and optimiser steps cut — an order-of-magnitude gap in training FLOPs is not an extraordinary claim. The architecture is consistent with the arithmetic.
Why it cannot be confirmed
No training compute figure, no cluster hours, no token count for the training corpus and no dataset disclosure accompanied the release. “One-ninth” is a ratio with no numerator published, so the claim is unfalsifiable in the strict sense. That is standard practice industry-wide and no better for being standard.
What it means if it holds
If a lab can reach this capability tier for a ninth of the previous generation’s training bill, the number of organisations that can train a competitive large language model rises sharply, and the strategic value of a compute moat falls just as sharply. That is a bigger story than any benchmark row in the table.
The efficiency argument in three ratios
Long Context and Speed in Qwen3.8-Flash
A million-token window is only useful if you can afford to fill it, which is where the reworked attention kernels come in.
The two context numbers
Native context is 262,144 tokens — 256K — and YaRN extension takes it to 1,000,000. The distinction matters: the native figure is what the model was trained to handle, and the extended figure is a rope-scaling trick applied at inference. Quality at 1M is a different question from quality at 256K, and Alibaba has not published a needle-in-a-haystack curve for either.
The speed claims
Alibaba reports prefill up to 7.6 times faster and decode up to 4.9 times faster at the 1M mark. The vLLM recipe page for the model quotes larger figures still — 10.2x prefill and 6.6x decode. Both sets are relative speedups against an unspecified baseline, and the discrepancy between them is unexplained.
Why the architecture delivers it
The mechanism is not mysterious. Three quarters of the layers use linear-time recurrence rather than quadratic attention, and the quarter that does attend is capped at a 2,048-token budget regardless of how long the context actually is. Attention cost stops scaling with sequence length, so the curve flattens.
What it means in practice
The practical consequence is that whole-repository prompting and long agent traces stop being a budget decision. For teams building AI agents that accumulate hundreds of turns of tool output, a flat attention cost changes the economics of context management more than the headline token price does.
Qwen3.8-Flash Pricing Against the Rest of the Lineup
The hosted rates are where the architectural efficiency turns into a commercial argument, and the comparison with Alibaba’s own flagship is stark.
The published rates
Qwen3.8-Flash will serve at $0.16 per million input tokens and $0.47 per million output tokens through QwenCloud. Those are the rates the Qwen team announced alongside the weights, for a production model that had not yet opened to traffic at the time of writing.
Against Qwen3.8-Max
Alibaba’s 2.4-trillion-parameter flagship is priced at $2.00 per million input tokens and $6.00 per million output tokens. That is 12.5 times the input rate and 12.8 times the output rate of Qwen3.8-Flash — the “roughly twelve times cheaper” figure that circulated on release day, and it is accurate.
The cost-per-task question
Cheaper per token is not automatically cheaper per task. A model that needs more reasoning tokens to reach the same answer can cost more at a lower rate. The benchmark scores argue that this one does not have that problem, but the only way to know for your workload is to run it. Our LLM API pricing comparison sets out how the rest of the market prices against these numbers.
Published rates per million tokens
Running Qwen3.8-Flash-Next on Your Own Hardware
This is the part where the open weights stop being an abstraction and start being a purchase order.
The checkpoint sizes
The FP8 checkpoint is 172.78 GiB. The BF16 checkpoint is 335.28 GiB. Neither fits on a single accelerator of any kind currently shipping, so tensor or expert parallelism is mandatory rather than optional.
Validated configurations
The vLLM recipe validates tensor-parallel-4 with the FP8 checkpoint on four GB300s, with TP2 as the minimum tested configuration at roughly 190 GiB per GPU. On eight H200s the recommendation is tensor-and-expert parallelism with the Triton MoE backend — plain TP8 is documented as incompatible with the FP8 checkpoint. Four MI355X cards are supported at TP4.
The host memory line
At least 51 GB of system RAM plus headroom is required if you offload the N-gram embedding table, which the recipe recommends. That is an unusual line item on a serving spec, and it is the direct consequence of the architecture: the capacity lives in a lookup table, and lookup tables live happily in cheap memory.
The quantised route
Unsloth’s GGUF ladder runs from 72.5 GB at UD-IQ1_S through 93.7 GB at UD-IQ4_XS to 111 GB at UD-Q4_K_XL. That puts a four-bit quantisation inside reach of a 128 GB unified-memory workstation, which is why Mac Studio owners were the most enthusiastic constituency on release day.
Who was disappointed
Owners of single 32 GB consumer cards were not. The most common complaint in the release-day threads was that Alibaba did not ship a smaller sibling — something in the 35B-A3B range that would sit comfortably in 32 GB — alongside the flagship preview. A one-bit quantisation of a 125B model is not what most desktop users were hoping for.
Hardware summary
| Route | Footprint | Practical target |
|---|---|---|
| BF16 checkpoint | 335.28 GiB | Multi-node or 8-way server |
| FP8 checkpoint | 172.78 GiB | 4 x GB300 at TP4, or 8 x H200 at TEP8 |
| FP8 at TP2 | ~190 GiB per GPU | Minimum validated configuration |
| UD-Q4_K_XL GGUF | 111 GB | 128 GB unified-memory workstation |
| UD-IQ4_XS GGUF | 93.7 GB | 128 GB workstation with headroom |
| UD-IQ1_S GGUF | 72.5 GB | 96 GB systems, heavy quality loss |
| N-gram table offload | 51 GB host RAM plus headroom | Recommended in every configuration |
The Licence Attached to Qwen3.8-Flash-Next
Open weights are not one thing, and the licence line on this model card is the detail most of the coverage skipped.
It is not Apache 2.0
The model ships under qwen-community-1.0. Much of the Qwen catalogue has historically been Apache 2.0, which is about as unencumbered as a licence gets. A bespoke community licence is a different instrument, and the difference is not cosmetic.
What a community licence typically changes
Vendor-specific community licences generally add some combination of attribution requirements, acceptable-use restrictions and a threshold above which separate commercial terms apply. Anyone planning production deployment needs to read the actual text rather than the word “open” in the headline.
Why it matters more here than usual
The whole commercial case for this model is self-hosting: run it inside your own boundary, avoid per-token billing, keep your source code and documents on your own infrastructure. Every one of those benefits depends on the licence permitting it at your scale. Our guide to open-weight AI models covers how the various licence families compare.
The pattern across Chinese labs
The contrast is worth noting. Z.ai released GLM-5.3-Flash under a plain MIT licence the same week, as we covered in the Ox Alpha reveal. Two near-frontier Chinese open-weight models, two very different legal instruments, released within days of each other.
Why Qwen3.8-Flash-Next Previews Qwen4
Shipping the architecture before the model family is a deliberate choice, and an unusual one.
The stated reason
Alibaba’s framing is that the community should be able to examine the architectural changes before Qwen4 is built on top of them. Kernels need writing, serving stacks need patching, quantisation schemes need adapting. Releasing the architecture early means all of that work is finished by the time the flagship arrives.
The unstated reason
There is a competitive dividend too. vLLM, SGLang and TokenSpeed all shipped support at or near launch. Every hour of engineering the open-source ecosystem spends on Gated DeltaNet, QSA and N-gram embedding lookups is an hour that makes Qwen4 easier to deploy than a rival architecture nobody has optimised yet.
The release cadence around it
This lands inside a fortnight that also produced GLM-5.3-Flash from Z.ai and a DeepSeek-V4-Flash comparison baseline. The Chinese open-weight cadence through 2026 has been relentless, and the pattern is consistent: publish weights, publish a technical report, price the hosted tier aggressively, and let adoption do the marketing.
What is still missing
Qwen4 itself has no announced date, no parameter count and no capability claims. Everything anyone knows about it is inferred from a preview model that Alibaba has been careful to describe as a preview. That is a real limit on how far the inference can be pushed.
What Qwen3.8-Flash Means for Technology Buyers
The practical response sits between the two obvious errors: adopting on the strength of a vendor table, or dismissing on the strength of a flag.
Evaluate on your own tasks
SWE-bench and CoWorkBench measure generic capability. They do not measure your codebase, your document formats or your review standards. A fifty-task internal evaluation against real tickets will tell you more in a week than the entire benchmark table will.
Model the total cost, not the token rate
At $0.16 and $0.47 per million tokens the hosted rate is close to a rounding error for most workloads. The costs that matter are integration, evaluation, monitoring and the engineering time to migrate. Those do not scale down with the token price.
Decide where inference runs before you decide what runs
The hosted API means data crossing into Chinese jurisdiction. The open weights mean it does not. For regulated sectors that single decision dominates every capability question, and it is answerable before any benchmark is read.
Treat the preview label seriously
Alibaba called this a preview of the Qwen4 architecture. Building a production dependency on a preview checkpoint, under a bespoke licence, from a lab that has said a successor architecture is coming, is a decision that needs an explicit exit plan attached to it.
Risks and Open Questions
Four things about this release remain genuinely unresolved, and the benchmark table settles none of them.
Nothing has been independently verified
Every number in this article comes from Alibaba, from the model card, or from serving documentation written against Alibaba’s claims. As of publication no independent evaluation of Qwen3.8-Flash-Next has been published. The weights are open, so that will change quickly — but it has not changed yet.
The training corpus is undisclosed
No dataset itemisation, no token count, no description of the post-training regime. The one-ninth cost claim rests on a training process that has not been described in enough detail for anyone to reason about contamination, licensing of training data, or how much of the efficiency comes from architecture versus data curation.
Long-context quality is unmeasured in public
A 1,000,000-token window obtained by YaRN extension from a 262,144-token native context is a specification, not a demonstration. Retrieval accuracy across that window has not been published, and long-context degradation is the failure mode that most commonly separates a claimed window from a usable one.
Preview status cuts both ways
The -Next suffix is a warning label. It signals that serving stacks may change, that quantisations may need redoing, and that the production Qwen3.8-Flash served through QwenCloud may not be weight-identical to the checkpoint you downloaded. Plan for divergence between the two.
Frequently Asked Questions
What is Qwen3.8-Flash-Next?
It is a multimodal mixture-of-experts model from Alibaba’s Qwen team, open-weighted on 26 August 2026. It has 125 billion parameters plus a 51-billion-parameter N-gram embedding table, activates 6 billion parameters per token, and is described by Alibaba as an early preview of the architecture Qwen4 will use.
How is Qwen3.8-Flash different from Qwen3.8-Flash-Next?
Qwen3.8-Flash-Next is the open-weight preview checkpoint you can download. Qwen3.8-Flash is the production model Alibaba will serve through the QwenCloud API at $0.16 per million input tokens and $0.47 per million output tokens. They are the same family; they are not guaranteed to be the same weights.
What makes the architecture new?
Four things: hybrid attention pairing Gated DeltaNet with Qwen Sparse Attention at block granularity, a Gated Residual stream widened to four gated branches, a 51-billion-parameter N-gram embedding table that can be offloaded to host RAM, and a training recipe using Muon and AdamW on different weight categories with batch-size warmup removed.
Does it really beat Claude Opus 4.6?
On the benchmarks Alibaba published, on Alibaba’s harness, it wins eight of nine — including +22.5 on AndroidWorld, +19.1 on JobBench and +9.1 on SWE-bench Pro — and loses Humanity’s Last Exam 35.9 to 40.0. No independent replication has been published yet, so the honest answer is that it appears to, and nobody outside the lab has confirmed it.
What hardware do I need to run it?
The FP8 checkpoint is 172.78 GiB and is validated on four GB300s at TP4 or eight H200s with tensor-and-expert parallelism; the minimum tested configuration is TP2 at about 190 GiB per GPU. A four-bit GGUF at 111 GB will run on a 128 GB unified-memory workstation. Add at least 51 GB of host RAM for the N-gram table.
How much cheaper is it than Qwen3.8-Max?
12.5 times cheaper on input and 12.8 times cheaper on output: $0.16 against $2.00 per million input tokens, and $0.47 against $6.00 per million output tokens.
Is it really open source?
It is open-weight rather than open-source. The licence is qwen-community-1.0, not Apache 2.0 or MIT, so read the terms before deploying commercially. The training data and training code were not released.
How long is the context window?
262,144 tokens natively, extensible to 1,000,000 tokens using YaRN. Alibaba reports prefill up to 7.6 times faster and decode up to 4.9 times faster at the one-million-token mark; the vLLM recipe quotes 10.2x and 6.6x against an unstated baseline.
Can it handle images and video?
Yes. Text, images and video are accepted as input, and text is returned. It scores 84.5 on AndroidWorld, which drives a real mobile interface, and 64.4 Pass@3 on ClawEval-MM.
When is Qwen4 coming?
Alibaba has not said. The only public information is that Qwen3.8-Flash-Next previews the architecture Qwen4 will be built on. No date, parameter count or capability claim has been published.
References and Further Reading
Qwen3.8-Flash-Next: A New Architecture, Towards Ultimate Efficiency
Qwen3.8-Flash-Next model card on Hugging Face
Qwen3.8-Flash-Next FP8 checkpoint
QwenLM/Qwen3.8-Flash-Next on GitHub
Alibaba releases Qwen3.8-Flash-Next, targeting “ultimate cost efficiency”
Alibaba releases Qwen 3.8 Flash-Next, beats Opus 4.6 on most benchmarks
Alibaba to release Qwen 3.8-Flash-Next as a preview of what Qwen 4 will offer
Qwen3.8-Flash-Next vLLM deployment recipe
Qwen3.8-Flash-Next SGLang cookbook entry
Unsloth GGUF quantisations of Qwen3.8-Flash-Next
Alibaba unveils new cost-efficient model Qwen3.8-Flash-Next
Qwen3.8-Flash-Next 125B-A6B: what the release means for local deployment
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.