Qwen 3.8 vs Kimi K3: Two Chinese Giants, and Now One Is Cheaper

These are the two most consequential Chinese frontier models of 2026, and they are close cousins: both enormous sparse Mixture-of-Experts systems, both at 1M-token context, both multimodal, both promising open weights. Kimi K3 from Moonshot is actually the larger of the two at 2.8T total parameters against Qwen's 2.4T — a useful reminder that parameter count is not a ranking.

Until August 3, 2026 this comparison was lopsided in a specific way: Kimi K3 had an audited Artificial Analysis Intelligence Index of 57.1, a #1 LMArena frontend-code ranking, published pricing, and shipped weights, while Qwen3.8-Max had a 2.4T headline and nothing else. GA changed the economics decisively. Qwen now publishes $2 / $6 per million tokens against Kimi's $3 / $15 — which makes the bigger, better-proven model the more expensive one by a wide margin.

A note for builders — the fastest way to settle this is to run both on your own prompts. OrcaRouter fronts 200+ models behind one OpenAI-compatible endpoint, so you can pit Qwen 3.8 Max against Kimi K3 on the same task without wiring up two SDKs.

TL;DR verdict. Kimi K3 still wins on proof: an audited AA Intelligence Index of 57.1 (#3), the LMArena frontend-code #1 spot at 1679, and open weights that are actually ready. Qwen3.8-Max remains unscored by every independent evaluator. But GA handed Qwen two real advantages. On price it is now substantially cheaper — $2 / $6 against $3 / $15, meaning 2.5× less on output. And in the one head-to-head third-party test that exists, Qwen lost the headline 80 to 83 while being the visibly better-behaved agent: 354 repo citations to 274, 22 gateway requests to 53, and zero failed tool calls to two. Kimi for audited quality; Qwen for cheaper, cleaner agentic execution.

Key takeaways

• Kimi K3 is the bigger model — 2.8T MoE against Qwen's 2.4T, roughly 400B more — which is a good argument against treating parameter count as a proxy for capability.

• Qwen3.8-Max is GA since August 3, 2026 at $2 / $6 per 1M flat, versus Kimi K3's $3 / $15 (AA blended ~$2.31). Qwen is now 2.5× cheaper on output.

• Kimi K3 holds the audited standing: AA Intelligence Index 57.1 (#3), behind only Fable 5 and GPT-5.6 Sol, plus LMArena frontend-code #1 (1679). Qwen3.8-Max is still unscored.

• In the only direct test (Trilogy AI), Kimi edged Qwen 83 to 80 overall — but Qwen made more repo citations (354 vs 274), fewer gateway requests (22 vs 53), and had zero failed tool calls (vs two), with a 9/10 tool-use rating to Kimi's 8/10.

• Qwen now reports agentic and coding numbers: Terminal-Bench 2.1 86.6, OSWorld-Verified 86.1, FrontierSWE 73.5 (from a predecessor's 40.7) — all Alibaba-reported.

• Openness timing still favors Kimi: its weights are essentially ready; Qwen's are promised within the week, with no license or repository yet.

Accuracy note: all Qwen3.8-Max quality figures are Alibaba-reported and unverified by third parties. Kimi K3 figures come from Artificial Analysis and LMArena. The head-to-head comparison is a single test by Trilogy AI and should be read as one data point, not a general ranking. Qwen pricing is the GA rate card and supersedes July's preview discount.

The specs and price, side by side

Here is the hard data, each figure attributed to its source.

• Maker / status — Qwen3.8-Max: Alibaba; GA (August 3, 2026); Kimi K3: Moonshot; open-weight release

• Architecture — Qwen3.8-Max: ~2.4T total, sparse MoE (active params still undisclosed); Kimi K3: 2.8T MoE, native vision (Artificial Analysis)

• Context window — Qwen3.8-Max: 1M tokens (983,616 with thinking; max output 131,072); Kimi K3: 1M tokens (Artificial Analysis)

• AA Intelligence Index — Qwen3.8-Max: Still unscored; Kimi K3: 57.1 — #3 (Artificial Analysis)

• Arena / leaderboard — Qwen3.8-Max: Not scored publicly; Kimi K3: Frontend-code #1, 1679 (LMArena)

• Vendor-reported scores — Qwen3.8-Max: Terminal-Bench 2.1 86.6, GPQA Diamond 92.6, OSWorld-Verified 86.1 (Alibaba-reported); Kimi K3: standing is third-party measured

• Pricing (in / out) — Qwen3.8-Max: $2.00 / $6.00 per 1M, flat; cached input $0.25; Kimi K3: $3 / $15 per 1M (AA blended ~$2.31), cache hits $0.30

• Open weights — Qwen3.8-Max: Promised within the week (no license yet); Kimi K3: Open-weight; weights ready

• Speed — Qwen3.8-Max: p50 TTFT 1.64s (OrcaRouter 7-day telemetry); Kimi K3: verbose and slow (~34s TTFT, per AA)

Two things frame this comparison, and one of them fully reversed on August 3.

First, the price relationship inverted. Through July, Kimi K3 was the model with a real price and Qwen was the model with a discount coupon. Now both publish rates, and the better-proven, larger model is the pricier one: $3 / $15 against $2 / $6. On output — where reasoning and code generation spend their money — Kimi costs 2.5× more. Qwen also charges a flat rate across its entire 1M-token context, and its cached input at $0.25 slightly undercuts Kimi's $0.30 cache-hit rate. For any high-volume workload, Qwen's economics are now clearly better.

Second, the proof gap is unchanged, and it is the reason Kimi still wins. Kimi K3's 57.1 Intelligence Index places it third overall, behind only Fable 5 and GPT-5.6 Sol — that is a neutral evaluator's verdict. Its LMArena frontend-code #1 at 1679 is a crowd-sourced result from many blind comparisons. Qwen's Terminal-Bench 86.6 and GPQA Diamond 92.6 are real numbers but Alibaba's own, on a harness nobody else has run. One useful anchor: the predecessor Qwen3.7-Max scored 46 on the AA index against Kimi's 57.1, so Qwen needs a large generational jump merely to draw level.

There is also a latency wrinkle worth noting, because it cuts against the usual story. Artificial Analysis measures Kimi K3 at roughly 34 seconds time-to-first-token — genuinely slow. OrcaRouter's GA telemetry shows Qwen3.8-Max at a p50 TTFT of 1.64 seconds. Those measurements come from different sources and are not a controlled comparison, but they complicate the preview-era assumption that Qwen is the sluggish one of the pair.

The one head-to-head test, read carefully

This is the only matchup in the series where a third party has run both models against each other on the same task, which makes it worth reading closely rather than summarizing to a winner.

Trilogy AI ran an architecture-understanding task — comprehend a real repository and reach a correct architectural conclusion — and scored the results:

• Overall score — Qwen3.8-Max: 80 / 100; Kimi K3: 83 / 100

• Repo citations — Qwen3.8-Max: 354; Kimi K3: 274

• Gateway requests — Qwen3.8-Max: 22; Kimi K3: 53

• Failed tool calls — Qwen3.8-Max: 0; Kimi K3: 2

• Tool-use rating — Qwen3.8-Max: 9 / 10; Kimi K3: 8 / 10

• Cache hit rate — Qwen3.8-Max: >90%; Kimi K3: >90%

Kimi won the headline by three points. But look at the process columns, because for agentic engineering they matter more than the score. Qwen reached the same core architectural conclusion using fewer than half the gateway requests while producing 29% more grounding citations and — the detail that should interest anyone building agents — zero failed tool calls against Kimi's two.

Failed tool calls are what actually break production agents. They cascade: a malformed call produces an error the model must interpret, which consumes turns, which risks further errors. A model that makes 22 clean requests where another makes 53 with two failures is cheaper and more predictable to operate, even if it scores three points lower on the answer. Combined with Qwen's 2.5× cheaper output rate, the operational economics of that run favor Qwen substantially.

The caution is that this is one task, one evaluator, one run. It is not a benchmark suite and it does not generalize. Kimi's 57.1 index and #1 frontend ranking rest on far more evidence than this single test does. The correct conclusion is narrow: on at least one realistic agentic task, Qwen was the more disciplined tool user while losing slightly on output quality.

Cost in practice: agentic runs at volume

Take a repository-analysis agent: 250,000 tokens of code context per run, 12,000 tokens of output.

• Qwen3.8-Max: 0.25M × $2 = $0.50, plus 0.012M × $6 = $0.072. ≈ $0.57 per run.

• Kimi K3: 0.25M × $3 = $0.75, plus 0.012M × $15 = $0.18. ≈ $0.93 per run.

• At 1,500 runs/day: Qwen ≈ $858/day; Kimi ≈ $1,395/day — roughly $16,000/month apart.

• With caching on a stable repo: Qwen's cached input at $0.25/1M takes the run to ≈ $0.14; Kimi's $0.30 cache-hit rate takes it to ≈ $0.26.

The gap is real at both ends, and the Trilogy result compounds it: if Qwen genuinely uses less than half the gateway requests to finish comparable work, the effective cost difference on agentic tasks is wider than the rate card alone suggests.

Both models bill thinking tokens as output, so reasoning-heavy configurations cost more than the naive estimate on either side — but Qwen's $6 rate makes that penalty 2.5× smaller.

Openness: near-parity, different timing

This is the axis where the two are most alike and the difference is purely about readiness. Kimi K3's weights are open and essentially ready. Qwen3.8-Max's are promised within the week, with no license text, model card, or repository yet — and a license is what determines whether you can use a model commercially at all.

Practically, neither flagship is self-hostable by a normal team. Qwen at 2.4T is roughly 1.2TB in 4-bit against about 141GB per H200; Kimi at 2.8T is larger still. Both need eight or more top-end accelerators, and Alibaba's undisclosed active-parameter count means you cannot even model Qwen's throughput.

Which is why the concurrently announced Qwen3.8-27B matters here. Also going open-weights and reportedly running in about 17GB of VRAM, it gives the Qwen family a genuine on-premise story that neither 2.4T nor 2.8T flagship provides. If open weights are your actual reason for choosing between these two, the 27B changes the calculus more than either giant does.

FAQ

Is Qwen 3.8 better than Kimi K3?

Not on audited evidence. Kimi K3 holds an independent AA Intelligence Index of 57.1 (#3) and LMArena's #1 frontend-code spot, while Qwen3.8-Max remains unscored by every third-party evaluator. In the one direct test available Kimi won 83 to 80. Qwen's advantages are price (2.5× cheaper output) and, in that same test, cleaner tool use.

Which model is bigger?

Kimi K3, at 2.8T total parameters against Qwen3.8-Max's 2.4T — roughly 400B more. Neither discloses enough for that to be meaningful, though: Alibaba has never published Qwen's active-parameter count, which is the figure that actually determines serving cost in a Mixture-of-Experts model.

Which is cheaper?

Qwen3.8-Max, clearly, since GA. It lists $2 / $6 per 1M against Kimi K3's $3 / $15 — 33% less on input and 2.5× less on output. Qwen's rate is flat across the whole 1M context, and its cached input at $0.25 slightly undercuts Kimi's $0.30 cache-hit rate.

What did the head-to-head test actually show?

Trilogy AI's architecture-understanding task scored Kimi 83 and Qwen 80. But Qwen produced 354 repo citations to Kimi's 274, used 22 gateway requests to 53, logged zero failed tool calls against two, and earned 9/10 on tool use to Kimi's 8/10. Kimi gave the better answer; Qwen was the more disciplined agent. It is one task, so treat it as a data point rather than a ranking.

Did Qwen 3.8 Max leave preview?

Yes, on August 3, 2026, with a published rate card and an OpenAI- and DashScope-compatible endpoint. July's Qoder credits campaign no longer describes how it is sold.

Which one is faster?

Possibly Qwen now, which reverses the preview-era assumption. Artificial Analysis measures Kimi K3 at roughly 34 seconds time-to-first-token; OrcaRouter's 7-day GA telemetry shows Qwen3.8-Max at a p50 TTFT of 1.64 seconds. Different sources and not a controlled comparison, but both models have a reputation for verbosity, and Kimi's measured TTFT is genuinely slow.

Can I self-host either?

Not realistically at flagship scale — 2.4T and 2.8T models need eight or more H200-class GPUs. Kimi's weights are ready today; Qwen's are promised within the week without a license yet. For a genuine on-premise option, the concurrently announced Qwen3.8-27B (reportedly ~17GB of VRAM) is the practical choice in the Qwen family.

Bottom line

Qwen 3.8 vs Kimi K3 is the closest matchup in this series, and general availability split it cleanly along two axes. Kimi K3 keeps the quality crown on the strength of things a referee measured: 57.1 on the Intelligence Index, third overall, and LMArena's top frontend-code slot. Those are not claims — they are results, and Qwen has none of the equivalent.

What Qwen won on August 3 is the economics, and it won them decisively: $2 / $6 against $3 / $15, flat across a million tokens, with slightly cheaper caching. Add the Trilogy finding that Qwen finished a comparable agentic task with half the gateway requests and zero failed tool calls, and Qwen looks like the more operationally efficient model even where it is the less accurate one. Watch two events: the first independent Intelligence Index score for Qwen3.8-Max, and the weights landing with a license. If Qwen scores anywhere near Kimi's 57.1, the price gap makes it very hard to justify Kimi for volume work. Until then, Kimi is the model you trust and Qwen is the one you can afford to run — and the honest way to choose is on your own repositories.