Kimi K3 Lands: Moonshot's 2.8-Trillion-Parameter Open Model Takes a Swing at the Frontier

TL;DR

On July 16, Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight Mixture-of-Experts model with a 1-million-token context window, and it went straight to the front page of Hacker News. Moonshot calls it the largest open-source model released to date and says its overall intelligence trails only Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol. On its own launch coding suite it goes further, taking first place outright on a couple of the hardest agentic tests. The pricing is the aggressive part: $3 per million input tokens, dropping to $0.30 on cache hits, and $15 per million output, with the full context window included at that rate. Open weights are promised imminently. The one honest caveat on day one is that the benchmarks and the pricing page arrived before the model card did.

The Headline Numbers

Two specs carry the story. The first is scale: per Moonshot's launch post, K3 is a Mixture-of-Experts model with 2.8 trillion total parameters (some earlier reports had pegged it nearer 2.5 trillion), which makes it the biggest set of open weights anyone has put on the table. The second is memory: a native 1-million-token context window (1,048,576 tokens, to be exact), which is enough to hold a small library of documents in a single prompt without chunking gymnastics.

Crucially, only 16 of its 896 experts, on the order of 50 billion parameters, are active on any given token. That gap between total size and active size is the whole trick, and it is what lets a model this large be priced like a much smaller one.

How a 2.8-Trillion-Parameter Model Stays This Cheap

A Mixture-of-Experts model is less one giant brain than a very large consulting firm. There are hundreds of specialists on the payroll, but any single question only pulls in the handful who are relevant, and only those bill their hours. K3 keeps 2.8 trillion parameters on staff and calls in just 16 of its 896 experts, on the order of 50 billion parameters, per token, so you get the knowledge of the whole firm at the running cost of a small team.

The long context leans on a stack of architecture changes Moonshot introduced with K3, led by Kimi Delta Attention, a linear-attention variant, plus Attention Residuals and a Stable LatentMoE design. Together they cut the memory the model has to hold while reading, and paired with a prefill cache that Moonshot reports hitting above 90% on coding workloads, they keep long-context inference cheap enough to price the way it does. That is the mechanism behind the native million-token window: not just a bigger buffer, but a cheaper way to attend across it. For scale, a million tokens is a working memory of roughly a dozen novels at once, which is a different category of task than "summarize this PDF."

What It Actually Scores

Then the launch benchmarks arrived, and they are more interesting than any single ranking suggests. On Moonshot's own published coding suite, with every model turned up to maximum thinking effort, K3 does not just place well, it wins some of the hardest agentic tests outright. It takes first on Program Bench (77.8, edging GPT-5.6 Sol's 77.6) and on SWE Marathon (42.0, ahead of Opus-4.8's 40.0), and finishes a razor-thin second on Terminal Bench 2.1 (88.3 to GPT-5.6 Sol's 88.8). It trails Fable 5 on the pure software-fixing tests, FrontierSWE (81.2 to 86.6) and Moonshot's internal Kimi Code Bench (72.9 to 76.9), and lands mid-pack on DeepSWE. Add it up and across six coding benchmarks K3 is first twice, second three times, and third once, measured against the strongest closed models on the market.

The honest framing: these are Moonshot's numbers, on Moonshot's chosen benchmarks, with the thinking dial cranked all the way up, which is the flattering configuration for any model. Independent aggregators are still catching up. BenchLM, the source the AI Bacon models page draws on, currently lists K3 as unranked until more non-generated coverage lands, though the cells it has logged (93.5 on GPQA-Diamond, 88.3 on Terminal Bench) line up with Moonshot's own. The coding results are genuinely good and the wins are real. Just remember that "beats Opus-4.8 on SWE Marathon" and "second only to Fable 5 overall" are sentences you want a few independent harnesses to co-sign before you rebuild your stack around them.

From K2 to K3 in About a Year

K3 did not appear from nowhere. Moonshot has been iterating in public at a brisk clip: Kimi K2 established the lab as a serious open-weight player in 2025, K2.6 was open-sourced in the spring, and the K2.7 line landed over the summer, with K2.7-Code reaching general availability in GitHub Copilot on July 1. K3 is the jump from a fast-moving point release to a new flagship architecture.

Why Builders Should Care

First, the price-to-capability ratio. If even the measured coding numbers hold up, K3 is one of the cheapest routes to near-frontier output, and $0.30 per million cached input tokens makes long-context, high-volume workloads that were painful on Western frontier pricing suddenly pencil out. PYMNTS framed the launch bluntly, as a bigger, cheaper challenge to Anthropic, and Moonshot has the momentum to make the framing stick: its annualized revenue reportedly crossed $300 million by mid-June against a roughly $20 billion valuation.

Second, open weights change the calculus. Once Moonshot posts the checkpoints (expected on its Hugging Face org, under what is reported to be a modified-MIT license), a frontier-adjacent model with a million-token context becomes something you can self-host, fine-tune, and run without sending a single token to anyone's API. That is a different value proposition than a closed model you merely rent.

Third, this is more pressure on the closed labs' pricing floor, not less. Every time an open-weight release lands within striking distance of the top and undercuts it on cost, the premium that OpenAI and Anthropic charge has to justify itself on capability alone. K3 is the loudest version of that argument so far.

Caveats

Plenty is still soft. There is no full model card with training details as of launch, the open weights are not out yet (Moonshot says they ship by July 27), and the modified-MIT license is expected rather than confirmed. The intelligence ranking and the 91.2 BrowseComp score are Moonshot's own claims, and BenchLM's refusal to rank the model yet is the tell: independent coverage is thin. As with any launch-day frontier claim from any lab, wait for third-party agent and SWE evaluations, and for the weights to actually download, before you treat the marketing as settled fact.

Key Takeaways

• Moonshot AI released Kimi K3 on July 16, 2026: an open-weight Mixture-of-Experts model with 2.8 trillion total parameters, 16 of 896 experts (about 50B) active per token, and a native 1-million-token context.
• Pricing is aggressive at $3 per million input tokens ($0.30 cached) and $15 per million output, with the full context included, positioning it as a bigger, cheaper alternative to the Western frontier.
• On Moonshot's own launch coding suite (all models at max thinking effort), K3 finishes first on Program Bench and SWE Marathon, second on Terminal Bench 2.1 and FrontierSWE, first twice and second three times across six benchmarks.
• Those are self-reported numbers in a flattering configuration; independent aggregator BenchLM still lists K3 as unranked pending more non-generated coverage, though its logged cells match Moonshot's.
• K3 caps a fast year (K2 to K2.6 to K2.7 to K3) and is the jump to a new flagship architecture built on Kimi Delta Attention, Attention Residuals, and a Stable LatentMoE design.
• Open weights are promised imminently under a reported modified-MIT license, but the model card, final specs, and third-party evals are not settled yet. Verify before you rebuild around it.

Sources: Kimi K3 launch post (Moonshot AI), BenchLM K3 benchmarks, BigGo Finance, OfficeChai, PYMNTS, GitHub Copilot changelog