Mercury 3 vs Kimi K3: One Model You Can Download, One You Can Only Rent

The most useful difference between Mercury 3 and Kimi K3 has nothing to do with speed or price. Kimi K3, which Moonshot AI released on July 16, 2026 and then published the weights for on July 27 under a Modified MIT licence, is a 2.8-trillion-parameter mixture-of-experts that you can download, inspect, quantise and host yourself. Mercury 3, which Inception released on October 19, 2026, is a closed diffusion model that you can only call. Everything else in this matchup — the 1M-token windows they share, the thousands of tokens per second one of them claims, the output price gap that runs in the opposite direction — follows from that first difference about who owns the artefact. The same key reaches the whole 200-plus-model catalogue at list price, with failover across providers. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

That framing is not academic. It decides which of these two models can survive a vendor going away, which one can be fine-tuned on your data, which one you can run at three in the morning without a rate limit, and which one is a line item that ends when someone else decides it does. It also decides what a fair comparison even is: Kimi K3's per-token price is a hosting cost, while Mercury 3's is a rental.

The specs, on both sides of the ownership line

• Parameters — Kimi K3 2.8 trillion total, 104 billion active per token, 16 of 896 experts vs Mercury 3 undisclosed, described only as more than 40× larger than Mercury 1

• Architecture — mixture-of-experts on Kimi Delta Attention vs a diffusion language model that refines blocks of tokens in parallel

• Context window — 1,048,576 tokens vs 1,000,000 tokens; effectively a tie

• Maximum output — Kimi K3 up to the full 1M window in a single response vs Mercury 3 131,072 tokens

• Input modality — text, image and video vs text only

• Weights — published, Modified MIT, on Hugging Face vs closed

• Input price — $3.00 per 1M vs $0.20 per 1M at list, $0.10 at launch

• Output price — $15.00 per 1M vs $0.70 per 1M at list, $0.35 at launch

• Cached input — $0.30 per 1M vs $0.02 per 1M

• Claimed throughput — not published by Moonshot vs 1,068 tokens per second, vendor-reported

• Independent Intelligence Index — 43.6, per Artificial Analysis vs no entry yet as of October 20, 2026

Put the two price columns next to each other and the gap is 15× on input and 21× on output. Put the modality line next to it and the gap narrows to something narrower than a rout: Kimi K3 accepts video and images, Mercury 3 accepts text, and a workload that has to watch a screen recording is not choosing between them at any price. Put the weights line next to it and the comparison changes character entirely, because the two prices are not measuring the same thing.

What Kimi K3's measurements actually say

Kimi K3 is the only model in this five-part series with both published weights and a full independent record, which makes it the useful anchor:

• Intelligence Index — 43.6, measured by Artificial Analysis

• Terminal-Bench 4.0 — 0.1263; Terminal-Bench 2.1 — 0.8502 under the vendor's own reported run

• SciCode — 0.5949, the highest of the four opponents in this batch

• Humanity's Last Exam — 0.4690

• AA-LCR long-context reasoning — 0.8867, the strongest long-context reading in the series

• GDPval-AA — Elo 1532.59

• AA-Omniscience — 19.7

• GPQA Diamond — 0.9354

• Cost per index task — $2.00

Two cautions about that table. The Terminal-Bench numbers are wildly apart — 0.1263 on the 4.0 version and 0.8502 on 2.1 — which is a benchmark-version problem rather than a model problem, and anyone quoting one without the other is picking a side. And the Artificial Analysis index for Kimi K3 reads 43.6 today, while the vendor's own launch material claimed 57 under an earlier revision of the index. The index has been rebuilt since; the 57 is a historical reading, not a current one.

The profile that survives both cautions is a model that is strong on long-context reasoning and scientific code, moderate on knowledge recall, and unremarkable on terminal-style agent work. That is not the profile of a model you download to replace a frontier API wholesale. It is the profile of a model worth owning for a specific slice of work — which is exactly the argument for open weights.

The download-and-run caveat, stated plainly

Published weights and runnable weights are different claims, and conflating them is the most common overclaim in coverage of large Chinese models. A 2.8-trillion-parameter mixture-of-experts with 104 billion active parameters is a datacenter artefact. Publishing the checkpoint means you may inspect it, fine-tune it, quantise it and serve it — for organisations with tens of accelerators and the engineering team to keep them busy. It does not mean it runs on a workstation, and the quantisation wave that follows every release of this kind is as much about making the model servable as making it small.

So the honest reading of Kimi K3's licence is narrower than "open" and still significant: it is an auditable, modifiable, self-hostable frontier-class model. For a regulated business that cannot send customer data to a third-party API, that is the whole ballgame. For a startup with three engineers and a $400 monthly API bill, it is a hobby.

Mercury 3 has no such option at any price. That is a real limitation and it is worth being explicit that Inception has not announced a weights release or a date for one. The enterprise alternative it does offer is contractual rather than technical: dedicated capacity, autoscaling, compliance controls and configurable data retention, all within Inception's platform. That answers the data-residency question for some buyers. It does not answer the "what if the vendor disappears" question for any of them.

Cost per completed job, not cost per token

The 21× output-price gap is the number that gets quoted, and on its own it is misleading in both directions.

First, Kimi K3's $3.00/$15.00 is a premium position in the Chinese market and Moonshot has said so. Against Mercury 3's $0.20/$0.70 it is not a premium, it is a different category. A team that runs 50 million output tokens a month pays $750 on Kimi K3 and $35 on Mercury 3 at list. That difference is small enough to be irrelevant for some businesses and large enough to fund a hire for others.

Second, the independent cost-per-task figures invert the intuition. Kimi K3 costs $2.00 per completed index task; Mercury 2.5, the closest measured member of the Mercury family, costs $0.1167 — but scored 12.3 against Kimi K3's 43.6. Cheap per token and cheap per unit of work are different numbers, and only one of them belongs in a capacity plan. Mercury 3 has no cost-per-task figure because it has no index score, which means the most important cell in this comparison is empty.

Third, and this is the one people forget: Kimi K3 is on OrcaRouter at Moonshot's list rate with 0% markup, and so is the rest of the 200-plus-model catalogue, on one OpenAI-compatible endpoint with automatic failover. Mercury 3 is not — it is served through Inception's own API and several third-party platforms. That asymmetry changes the shape of a sensible trial: you can compare Kimi K3 against your incumbent without signing anything new, and you cannot do the same for Mercury 3 without a second integration.

For a workload that runs long agent sessions, there is one more cost lever worth checking before either rate card is trusted. Kimi K3's 1M context has no published long-request surcharge, and neither does Mercury 3's. That is unusual and it is the main reason both models are credible for transcripts that would trigger a repricing tier elsewhere.

Which one to pick, by situation rather than by score

Choose Kimi K3 when the work needs images or video, when long-context reasoning is the core task — its 0.8867 on AA-LCR is the best reading in this batch — when you need to fine-tune on proprietary data, when data cannot leave your infrastructure, or when you want a measured capability floor before committing traffic. The weights are the strategic asset; the licence is genuinely permissive; the price is the cost of not running it yourself.

Choose Mercury 3 when the workload is text-only, latency-bound and high-volume: support replies, search answers, tutoring turns, classification-adjacent work where a sub-second response is the product. At $0.70 output and $0.02 cached input it is cheap enough that per-request economics stop being a design constraint, and the 1M window means you rarely summarise. What you give up is modality, a published capability floor, and any path to self-hosting.

Choose both when the split is by lane rather than by model. The two systems are architecturally opposite — one is a massive sparse autoregressive MoE, the other is a parallel diffusion generator — which means they fail differently, and a policy that sends vision and long-context reasoning to Kimi K3 and instant text turns to Mercury 3 is coherent on the merits rather than on price alone. The catch is integration: only one of them is a single-key addition.

What would change the answer

Three developments, none of them speculative:

• Mercury 3's first independent score. The Mercury family's only measured member sits at 12.3 on the Intelligence Index, against Kimi K3's 43.6. If Mercury 3 lands near Kimi K3 while costing a fifteenth as much, the ownership argument becomes the only argument Kimi K3 has left. If it lands near 12.3, the price gap is irrelevant

• A Mercury 3 weights release. Inception has said nothing about one. Open weights would collapse the central difference in this article overnight

• A cheaper host for Kimi K3. The $3/$15 rate is Moonshot's own; the open licence means third-party hosts can undercut it, and a hosted 2.8T MoE at half the price changes the cost-per-task maths without changing the model

Until one of those lands, the comparison is between a rentable, unmeasured, very fast text model and an ownable, measured, premium model that reads video. Those are not substitutes. The only genuinely wrong answer is picking one because it wins a column that the other has never been scored in.