Qwen 3.8 vs Kimi K3: Specs, Benchmarks, and Which One You Can Actually Use (2026)

For teams deciding between Alibaba's newest flagship and Moonshot's open-weights giant.

Two frontier-scale Chinese models landed within eight days of each other in July 2026, and they represent opposite bets. Moonshot released Kimi K3's full weights on July 27 under its own custom Kimi K3 License, with published API pricing and third-party benchmark rankings already on the board. Alibaba previewed Qwen 3.8-Max on July 19 with a 2.4 trillion parameter headline and a claim that it sits "second only to Claude Fable 5," then launched it on August 3 with a live API and published pricing, but still no published benchmarks and no weights.

That difference matters more than the parameter counts. Right now, one of these models can be independently evaluated and self-hosted. The other can be priced and built on, but has to be taken on Alibaba's word until benchmarks or weights ship.

TL;DR comparison

What Qwen 3.8-Max is

Alibaba previewed Qwen 3.8-Max at the World AI Conference in Shanghai on July 19, 2026, then launched it on August 3 with a generally available API. It is a 2.4 trillion parameter sparse mixture-of-experts model, the largest Qwen to date, and Alibaba's internal evaluations place it second only to Claude Fable 5. Those evaluations have still not been published, so there are no scores, no methodology, and no benchmark names to check.

What is confirmed: roughly 95B active parameters per token, about 4 percent of the total, a 1M token context window with up to 128k output tokens, text and visual inputs, and API compatibility with both the OpenAI and Anthropic specs. The rate card is published: $2 per million input tokens, $6 per million output, $0.25 for cached input. The Token Plan subscription and the Qoder platforms remain the individual-use route. Broader modality claims are still unconfirmed. We track the full picture in our Qwen 3.8-Max release post.

The open-weights question is the big one. At launch Alibaba committed to releasing weights within about a week, and not just for the flagship: a smaller Qwen3.8-27B is coming open-weight alongside it. No license has been named. Until the weights actually land, Qwen 3.8 is an API-only model.

What Kimi K3 is

Kimi K3 is Moonshot AI's 2.8 trillion parameter model built on their LatentMoE framework, activating 16 of 896 experts per token (104B active parameters), with a 1 million token context window and native vision. The architecture combines Kimi Delta Attention with Attention Residuals, and Moonshot claims roughly 2.5x better scaling efficiency than Kimi K2. That efficiency number is vendor-published, so treat it as a claim, not a measurement.

The weights shipped on July 27, 2026 in MXFP4 format, 1.56 TB of them on Hugging Face, under Moonshot's custom Kimi K3 License. Moonshot itself calls the model open weight rather than open source, which is the honest label. Moonshot also contributed a vLLM implementation, so the serving path exists on day one. Full specs and the hardware math are in our Kimi K3 specs and hardware requirements post.

The benchmark situation: rankings vs promises

This is the clearest difference between the two models, and it is not close.

Kimi K3 is already on the independent leaderboards. It ranks second on the Vals AI index and sits in the top five of Artificial Analysis's Intelligence Index, where positions shift as new models get indexed. Community tracking has placed it first in Frontend Code Arena. Several outlets have called it the strongest open-weights model released to date.

Qwen 3.8-Max is not indexed anywhere yet. The "second only to Fable 5" line comes from Alibaba's internal testing, with nothing published to support it. One early independent head-to-head we covered scored K3 at 83 and the Qwen 3.8-Max preview at 80 on a real-world architecture evaluation, which is a single data point, not a verdict. The verified baseline for the Qwen family is still Qwen 3.7-Max: 92.4 on GPQA Diamond, 80.4 on SWE-bench Verified, 91.6 on LiveCodeBench.

We break down every verified and unverified claim in Qwen 3.8 Benchmarks: What's Actually Verified So Far. The short version: if your procurement process requires evidence, only one of these models has any.

Pricing

Kimi K3's API pricing is public: $3.00 per million input tokens, $0.30 for cache hits, and $15.00 per million output tokens. Moonshot claims cache hit rates above 90% on coding workloads, which would pull effective input cost down substantially. That claim is vendor-published, so validate it on your own traffic before building a budget around it.

Qwen 3.8-Max's rate card arrived at launch: $2.00 per million input tokens, $6.00 per million output, $0.25 for cached input. On list price it undercuts K3 on both sides, and the output gap is 2.5x. What pricing can't tell you yet is cost per solved task, because Qwen 3.8 has no verified benchmarks. A cheaper token that needs more attempts isn't cheaper. Here's the full picture of Qwen 3.8 API access today, from the standard API to Token Plan tiers.

Self-hosting: possible vs not possible

Kimi K3 can be self-hosted, with a serious caveat: the weights are 1.56 TB, and Moonshot recommends 64 or more accelerators for production serving. That is multi-node territory, not a single 8-GPU box. The vLLM path exists, but this is a cluster-scale deployment with cluster-scale networking and orchestration problems. We walk through the full memory math in our Kimi K3 hardware requirements breakdown.

Qwen 3.8-Max cannot be self-hosted at all until weights ship. That may change soon. At launch Alibaba committed to open weights within about a week, and the more interesting piece may be the smaller Qwen3.8-27B coming with them. K3 self-hosting starts at 1.56 TB and cluster scale; a 27B-class model runs on a single rented GPU. If both drops happen as promised, Qwen would cover both ends of the self-hosting spectrum while K3 remains cluster-only.

Choose Qwen 3.8-Max if

• You want the newest flagship through a managed API at aggressive list pricing ($2 in / $6 out)
• You want OpenAI- or Anthropic-compatible endpoints and can tolerate unverified benchmarks and a single provider for now
• You plan to self-host later: the promised 27B open-weight release would be far more practical to run than K3's 1.56 TB

Choose Kimi K3 if

• You need a frontier-scale model you can evaluate against third-party data today
• Published, predictable API pricing matters for your cost model
• Open weights matter, whether for auditability, fine-tuning plans, or avoiding API lock-in

Frequently asked questions

Is Qwen 3.8 better than Kimi K3? Unknown. Alibaba claims Qwen 3.8-Max is second only to Claude Fable 5, but has published no benchmarks. Kimi K3 has third-party rankings and one early head-to-head favored it. Until Qwen 3.8 is independently indexed, there is no evidence-based answer.

Is Kimi K3 open source? No, and Moonshot doesn't claim it is. The weights are open, released July 27, 2026 under a custom Kimi K3 License, and Moonshot describes the model as open weight. Training data and code are not open.

Is Qwen 3.8 open source? Not yet. At launch Alibaba committed to open weights within about a week, for both Qwen 3.8-Max and a smaller Qwen3.8-27B. No license has been named.

What does the Kimi K3 API cost? $3.00 per million input tokens ($0.30 on cache hits) and $15.00 per million output tokens, per Moonshot's published pricing.

Can I run Kimi K3 on my own GPUs? Only at cluster scale. The weights are 1.56 TB and Moonshot recommends 64+ accelerators for production serving. There is a vLLM implementation available at release.

What is the cheapest way to try both? Kimi K3 through Moonshot's API. Qwen 3.8-Max through Alibaba's standard API at $2 in / $6 out per million tokens, or the Token Plan subscription for individual use. Neither requires infrastructure to evaluate.

When will Qwen 3.8 weights be released? Alibaba committed to a release within about a week of the August 3 launch. As of this update they have not landed on Hugging Face.

Bottom line

Kimi K3 is still the model you can evaluate on evidence: published pricing, open weights, third-party rankings, and a real self-hosting path if you have the cluster for it. Qwen 3.8-Max closed most of its gap at launch: a live API, aggressive published pricing, and disclosed specs. What it still doesn't have is proof. Until an official benchmark table or the promised weights arrive, K3 is the known quantity and Qwen 3.8-Max is the cheaper bet with the bigger claim.

There's also a third door: an open-weights frontier model with published numbers you can check today, at a size that doesn't need a 64-accelerator cluster. That contrast is our Qwen 3.8 vs GLM 5.2 comparison.

If you are benchmarking either one, our breakdown of what's actually verified about Qwen 3.8 is the place to start. And if you want to run frontier open models without owning the hardware problem, take a look at Yotta's AI Gateway, where models like Qwen3.7-Max are live behind one OpenAI-compatible API, or talk to us about multi-node GPU capacity for K3-scale deployments.