Open Weight Models Are No Longer Just Catching Up

For the past three years, open-weight model development followed the same pattern. A closed lab sets the frontier, and six to nine months later an open-weight model reproduces most of that capability at a lower price. This made open models useful, but left their users trailing behind the state-of-the-art. If you wanted the strongest available model for production work, you paid the closed-model tax and accepted whatever terms came with it.

Kimi K3, the open-weight model released yesterday by Moonshot AI, breaks this pattern. It is a Mixture-of-Experts model with 2.8 trillion total parameters and a native 1M-token context window, and on the benchmarks that measure sustained agentic work, it now sets the frontier rather than chasing it.

Below, we look at what that means for anyone running agents in production, then explore two of its main architectural decisions, Kimi Delta Attention and Attention Residuals.

Benchmarks and implications for agent design

All models evaluated at maximum thinking effort. K3 is the only open-weight model in the top tier.

K3 is at the frontier across the board: first or second on five of six coding benchmarks, performing at the same level as Fable 5, GPT-5.6 Sol, and Opus 4.8 rather than a tier below them. On SWE Marathon, which measures multi-hour agentic work, K3 leads outright at 42.0, while the previous generation of open-weight models sat near 14.

For developers, the headline is simple: frontier-level agentic capability is now available as open weights, not only through closed APIs. Teams running Claude or Codex for agentic coding can now achieve competitive quality on the same work while retaining full operational control: no rate limits you do not set yourself, no deprecation schedule you do not control, no access that gets blocked, and no terms that change under you. The model can be deployed inside your own infrastructure, so prompts, code, and agent traces never leave your environment.

In practice, an agent built on K3 behaves comparably to one built on the leading closed models. It stays on task across hundreds of steps, recovers from failed tool calls instead of spiraling, and keeps working when the trajectory runs long. That long-horizon reliability is what separated Claude- and Codex-based agents from everything else, and it is exactly where earlier open-weight generations fell short.

Implications for cost

K3 launched at $3 per million input tokens and $15 per million output tokens, with the full 1M context included. That is Sonnet pricing: Anthropic's mid tier sits at the same $3 and $15. What changed is what those dollars buy. The frontier tier, Opus 4.8 at $5 and $25 and GPT-5.6 Sol at $5 and $30, no longer has the quality to itself; K3 now sits at that level on the agentic benchmarks. You now get Opus-level quality at Sonnet pricing.

Over the past two years, closed-lab prices have barely moved: Opus has held at $5/$25 across four generations, and OpenAI's frontier models have stayed in the same band. That premium was defensible while open-weight models sat a tier below on quality, and the cheap rows in the chart, K2.7 and GLM-5.2, were the trade you made for it. K3 removes the trade: the quality gap has closed, but the price gap has not.

And the price gap will most likely widen, because the two sides face different market structures. A closed model has exactly one seller, so its price is set to recover training cost and margin, and it moves only when the lab decides it should. An open-weight model has as many sellers as care to host it. Dozens of providers will serve K3 within days, undercutting each other toward marginal serving cost; K2.7's cheapest hosted rate fell more than 20% in its first two months through exactly this dynamic. The launch price is where competition starts, not where it ends. This is the part frontier closed labs cannot follow.

For agent workloads the arithmetic is straightforward. Long trajectories are mostly decode, so the output token rate is the one that matters most, and K3's is 40% under Opus, while Kimi beats Opus on SWE Marathon. We believe dollars per completed task is a superior metric to dollars per token, and on this benchmark that measures completed long-horizon tasks, the winning model is also the cheaper one. Also, for a closed model, the API price is a floor, while for K3 it is a ceiling: any team with sustained volume can serve it on their own hardware and land under the published rate. It's important to keep in mind that cost per task depends on token efficiency as much as token price.

Under the hood

Two model architecture decisions carry the release. The first is Kimi Delta Attention (arXiv:2510.26692), a linear attention mechanism with per-dimension forget gates, interleaved with full attention at a 3:1 ratio. It is the first linear attention design to outperform full attention in matched comparisons, and it is what makes the 1M-token window economical to serve: KV cache footprint drops by up to 75%, and decoding at 1M-token context runs up to 6x faster than full attention. KV cache is what caps batch size in agentic serving, so this translates to several times more concurrent long-context sessions per GPU. It is the difference between a million-token agent you can demo and one you can run all day.

The second is Attention Residuals (arXiv:2603.15031). Every modern LLM sums layer output uniformly with residual connection. AttnRes replaces that fixed sum with learned softmax attention over depth, analogous to the shift from fixed aggregation to learned attention that sequence modeling underwent years ago. The result is a model that matches a baseline trained with 1.25x more compute, at under 2% added inference latency, with the biggest gains in multi-step reasoning and code. The capability arrives at training time and costs almost nothing to serve.

Both mechanisms were co-designed with Kimi kernels and inference paths from the start, leading to the model's strong benchmark results.

The catch-up era is over

With K3, the model is no longer the constraint. What defines the economics and the performance of agentic systems from here is the rest of the stack: the serving engine that exploits what these architectures expose, the fixed-size attention state, the hybrid cache layout, the block-level residual structure, instead of treating the model as a black box.

That is what we are building at LithosAI: the fastest inference engine for agentic workloads.

Kimi K3 is coming to LithosAI when weights drop!

Until then, here is our engine running Kimi K2.7 Code at over 1,000 peak tokens per second per user.