Kimi K3 vs Qwen 3.8: What's Confirmed, What Isn't (July 2026)

Kimi K3 vs Qwen 3.8: What's Confirmed, What Isn't

Three days after Moonshot AI's Kimi K3 launched, Alibaba's Qwen team previewed Qwen3.8-Max-Preview — a 2.4-trillion-parameter model it describes as matching frontier performance and "second only to Fable 5." The timing is not a coincidence: Qwen 3.8 arrived directly into the news cycle Kimi K3 created, and Alibaba's own framing positions it as a direct answer to K3's momentum.

Here's the catch for anyone trying to actually compare the two: Qwen 3.8 does not have a single independently verified benchmark published as of this writing. Alibaba's "second only to Fable 5" claim is company positioning, not a leaderboard result. This guide lays out exactly what's confirmed on each model, what's still just a claim, and what that means if you have to pick one today.

Side-by-side comparison illustration of Kimi K3 and Qwen 3.8 with a confirmed-vs-unconfirmed data panel and parameter-count bars

Quick answer

Kimi K3 has a real, testable track record right now: published pricing, an independently-run #1 leaderboard result (LMArena Frontend Code Arena), and third-party benchmark data from Artificial Analysis.
Qwen 3.8 is a preview with almost nothing independently verified yet: no published context window, no per-token API pricing, and zero third-party benchmark scores. Alibaba's performance claims are unconfirmed marketing language.
If you need a model you can evaluate and deploy today, Kimi K3 is the only one of the two with enough public information to do that.
If you're just monitoring the space, Qwen 3.8 is worth watching once Alibaba publishes real numbers — but there's nothing to test yet.
Neither has downloadable open weights yet. K3's are expected July 27, 2026; Qwen 3.8's are promised "soon" with no date.

Why this comparison is trending — and why most versions of it are wrong

"Qwen 3.8 vs Kimi K3" is one of the fastest-rising AI search terms this week, which makes sense: two of the largest open-weight models ever announced launched three days apart. But because Qwen 3.8 is still a preview with no public benchmarks, any article claiming to show you head-to-head scores between the two is either quoting unverified vendor marketing as fact or inventing numbers that don't exist yet. We're not going to do that. Below is what each company has actually confirmed, sourced to primary coverage, with the gaps left visibly open rather than papered over.

Release timing and positioning

	Kimi K3	Qwen 3.8 (Qwen3.8-Max-Preview)
Vendor	Moonshot AI	Alibaba (Qwen team)
Announced	July 16, 2026	July 19, 2026
Status	Generally available via API, app, and OpenRouter	Preview only, via Token Plan / Qoder / QoderWork
Positioning	Open-weight frontier generalist, agent-optimized	Multimodal flagship, positioned as chasing Kimi K3's momentum
Vendor's own framing	"Largest open-weight model released to date"	"Matches frontier models, trails only Fable 5" (unverified)

Qwen 3.8's release reads as a direct response to Kimi K3: Alibaba had kept its Max-tier flagships closed and API-only through Qwen3.7-Max (released May 2026), and Qwen 3.8 is the first of that line Alibaba has said will go open-weight — the same move Moonshot made with K3, days after K3 dominated the news cycle.

What's actually confirmed on each model

	Kimi K3	Qwen 3.8
Total parameters	2.8 trillion	2.4 trillion (headline figure; not independently verified)
Active parameters / MoE config	16 of 896 experts active per token	Not disclosed
Context window	1,048,576 tokens (~1M)	Not published
Multimodal input	Native image understanding	Images, video, and documents (Alibaba's first >1T-parameter multimodal model)
API pricing	$3.00 / MTok input, $15.00 / MTok output, $0.30 / MTok cached	Not published — preview access priced at "10% of standard pricing" through Token Plan, with no standalone rate disclosed
Independent benchmark data	Yes — Artificial Analysis Intelligence Index (57), LMArena Frontend Code Arena (#1, independently crowd-voted)	None published
Open weights	Promised, expected July 27, 2026	Promised "soon," no date, no Hugging Face repo yet
License	Modified MIT-style (planned)	Unconfirmed

The gap in that table is the whole story right now. Kimi K3 has a price you can put into a spend model and a leaderboard result run by a third party. Qwen 3.8 has a parameter count and a sentence of marketing copy.

The one real signal from Qwen 3.8: what it says about Moonshot

Even without benchmarks, Qwen 3.8's launch is informative for a different reason: it's a data point on how seriously competitors are taking Kimi K3. Moonshot AI reportedly crossed $300 million in annual recurring revenue in June 2026 and is preparing an IPO, according to Bloomberg reporting cited alongside Qwen 3.8's announcement. Alibaba previewing a 2.4T open-weight flagship — breaking from its past pattern of keeping Max-tier models closed — three days after K3's launch is a competitive reaction, not a coincidence. If you're trying to read the market rather than pick a model today, that's the more reliable signal than either company's benchmark claims.

Kimi K3's independently verified track record

Since Qwen 3.8 doesn't have comparable data yet, here's what's actually been checked on Kimi K3 by parties other than Moonshot:

LMArena Frontend Code Arena — a crowd-voted, independently-run leaderboard, not a vendor benchmark. K3 went from #18 to #1 within hours of launch, ahead of Claude Fable 5, and placed first in 6 of 7 frontend domains.
Artificial Analysis — an independent evaluator — scored K3 57 on its Intelligence Index (v4.1), ahead of Claude Opus 4.8 (56) and behind Claude Fable 5 (60), and separately measured a private long-horizon knowledge-work Elo of 1,547 for K3, a jump of +732 points over Kimi K2.6.
Cost-efficiency: Artificial Analysis put K3's cost-per-task at roughly $0.94, slightly below GPT-5.6 Sol's $1.04 on the same private evaluation.

None of this is a guarantee K3 is "better" than whatever Qwen 3.8 turns out to score once it has real numbers — but it's the difference between evaluatable and not-yet-evaluatable.

A near-term complication: Kimi K3 is capacity-constrained right now

Worth knowing if you're deciding which one to actually try this week: Moonshot paused new Kimi K3 subscriptions on July 19, 2026, after demand pushed close to the limits of its GPU capacity within 48 hours of launch. Existing subscribers and API customers are unaffected, and Moonshot is reopening new signups in batches — full details in why Kimi K3 sold out. If you're trying to get hands-on access to a new account specifically through the consumer app this week, that's a real practical obstacle Qwen 3.8 doesn't currently have, simply because Qwen 3.8's preview program is smaller and newer.

Which one should you use right now

Use Kimi K3 if:

You need a model you can actually evaluate today with real pricing and independently-verified benchmark data.
Your workload is agentic coding, frontend generation, or long-horizon tool use — the categories where K3's independently-verified results are strongest.
You can work within (or wait out) the current subscription pause, or you're accessing it through the still-open API and OpenRouter.

Watch Qwen 3.8, but don't commit yet, if:

You want a second open-weight option from a different vendor for redundancy — but you can't build a real evaluation plan around it until Alibaba publishes pricing, context window, and benchmark data.
Multimodal document/video handling at Alibaba's scale is specifically valuable to you and you're willing to wait for the open-weight release and real numbers.

Don't trust any article claiming hard head-to-head benchmark numbers between these two models today — if Qwen 3.8 hasn't published them, that content is either citing unverified vendor claims as fact or fabricating figures.

What to check back for

Qwen 3.8 benchmark scores from Artificial Analysis or a comparable independent evaluator.
Qwen 3.8's official per-token API pricing outside the promotional Token Plan preview rate.
Qwen 3.8's context window and active-parameter/MoE configuration.
Both models' open-weight releases — K3's is expected July 27, 2026; Qwen 3.8's has no confirmed date yet.

We'll update this comparison with real numbers as Alibaba publishes them, rather than backfilling estimates now.

FAQ

Is Qwen 3.8 better than Kimi K3? Unknown as of this writing. Alibaba claims Qwen 3.8 is "second only to Fable 5," but has not published any independent or self-reported benchmark scores to support that. Kimi K3 has independently-verified results (LMArena Frontend Code Arena, Artificial Analysis) that Qwen 3.8 currently does not.

Is Qwen 3.8 bigger than Kimi K3? No — Qwen 3.8 is reported at 2.4 trillion total parameters, versus Kimi K3's 2.8 trillion. K3 remains the larger of the two by the headline parameter count.

Is Qwen 3.8 open source? Alibaba has said open weights are coming "soon," but as of this writing there's no Hugging Face repo, no confirmed release date, and no confirmed license for Qwen 3.8. Kimi K3 is in a similar pending state, but with a specific committed date: July 27, 2026.

How much does Qwen 3.8 cost to use? Alibaba hasn't published standalone per-token API pricing for Qwen 3.8. Preview access is available through Alibaba's Token Plan, Qoder, and QoderWork at "10% of standard pricing," but that's a promotional preview rate, not a published production price.

Why did Alibaba release Qwen 3.8 right after Kimi K3? The timing — three days after K3's launch — plus Alibaba's explicit "second only to Fable 5" framing and its break from keeping Max-tier models closed, reads as a direct competitive response to Kimi K3's launch momentum and Moonshot AI's reported growth (roughly $300M ARR as of June 2026) and IPO preparations.

Can I use Qwen 3.8 right now? Only in preview, through Alibaba's Token Plan, Qoder, or QoderWork. There's no general API access or open-weight download yet.

Kimi K3 vs Qwen 3.8: What's Confirmed, What Isn't (July 2026)