Skip to content
NewsModel release

Moonshot AI releases Kimi K3, the world's first open 3T-class model at 2.8 trillion parameters

· by Pondero Newsdesk

The short version

Kimi K3 from China's Moonshot AI is the world's first open 3T-class model: 2.8 trillion parameters, 1-million-token context, and benchmark scores above Claude Opus 4.8 on most coding tasks. Open weights publish July 27.

Moonshot AI releases Kimi K3, the world's first open 3T-class model at 2.8 trillion parameters

Moonshot AI shipped Kimi K3 on July 16, 2026: a 2.8-trillion-parameter mixture-of-experts model that outscores Claude Opus 4.8 on most coding and agentic benchmarks while pricing at $3 per million cache-miss input tokens. Full open weights land July 27, which will make the model self-hostable for the first time.

What

Kimi K3 uses a sparse mixture-of-experts design with 896 total expert subnetworks, activating 16 per forward pass per the Kimi K3 announcement. Two new architectural components drive the scale: Kimi Delta Attention (KDA), which restructures long-sequence information flow, and Attention Residuals (AttnRes), which selectively retrieves representations across model depth. A Stable LatentMoE training framework keeps expert routing stable at 2.8 trillion parameters, where earlier MoE approaches have struggled. Moonshot says the design delivers roughly 2.5 times the scaling efficiency of Kimi K2.

The 1-million-token context window ships with a prefix-caching scheme Moonshot contributed to the vLLM project. Per the announcement, K3 achieves a cache-hit rate above 90 percent on coding workloads, which drops the effective input cost well below the $3/MTok list price. Output tokens cost $15/MTok; cache-hit input drops to $0.30/MTok.

On benchmarks reported by Moonshot, K3 outperforms Claude Opus 4.8 on five of seven coding measures: DeepSWE (67.5 vs. 59.0), FrontierSWE (81.2 vs. 66.7), SWE Marathon (42.0 vs. 40.0), Terminal Bench 2.1 (88.3 vs. 84.6), and BrowseComp (91.2 vs. 84.3). Anthropic's Fable 5 leads K3 on most of those same benchmarks. Moonshot's own announcement states that K3's "overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol."

K3 is available today on kimi.com, the Kimi Work desktop app (version 3.1.0 or later), and the Kimi Code terminal agent. Vision input (images and video) is supported natively. Max reasoning effort is on by default at launch; low and high effort modes are planned for later updates.

Why it matters

The raw parameter count sets a new ceiling. At 2.8 trillion total parameters, K3 is roughly 75 percent larger than DeepSeek V4 Pro (approximately 1.6 trillion parameters, per VentureBeat), which held the prior open-weight size record. Per Moonshot, for nine of the past twelve months Kimi models set the upper bound on open-model sizes. Extending that streak to 3T-class territory means developers now have a self-hostable model that beats a current frontier closed model on several standard coding evals.

The pricing angle is the more actionable point for operators. Frontier closed models that match Opus 4.8's coding scores typically carry significantly higher API costs. K3 at $3/MTok cache-miss input (and an effective input cost closer to $0.30/MTok at high cache-hit rates in coding) is a material change to the cost curve for teams running large agentic coding pipelines. Self-hosted deployments will trade GPU hosting costs against API spend once the July 27 weights arrive, a calculation that favors self-hosting for teams with spare H200 or similar capacity.

The benchmark picture requires a caveat. K3 does not beat Opus 4.8 uniformly. On HLE-Full, Opus 4.8 scores 49.8 against K3's 43.5 per Moonshot's benchmark table. On GPQA-Diamond, K3 leads (93.5 vs. 91.0), but GPT 5.6 Sol tops both (94.1). Teams evaluating K3 for agentic workloads should run task-specific evaluations rather than rely on aggregate comparisons. Moonshot itself recommends using a compatible harness such as Kimi Code and notes that switching to K3 mid-session from another model can destabilize generation.

Context and reactions

The release timed to land ahead of the 2026 World Artificial Intelligence Conference in Shanghai, according to VentureBeat. Moonshot AI is backed by Alibaba and is Beijing-based. VentureBeat noted the release also marks what the company described as a comeback: its market position had eroded significantly over the past 18 months following DeepSeek's rapid rise.

Moonshot's in-house evaluation of K3 showed the model capable of autonomous, long-horizon technical work, including a 48-hour chip design run using open EDA tools that produced a working 1.46 million standard-cell layout, and a GPU compiler (MiniTriton) that matched Triton's performance on several benchmarks. Those demonstrations are reproducible only once weights and the technical report are public, but they set specific, falsifiable claims rather than vague capability assertions.

The architecture's vLLM contribution is a concrete upstream signal. Moonshot pushed a KDA-compatible prefix-cache implementation to the vLLM community, timed to coincide with the K3 release. That move accelerates third-party inference support before the July 27 weight drop.

What to watch next

The July 27 open-weight release is the first hard milestone. Once weights land, independent benchmark runners typically produce verified results within 24 to 48 hours, and community fine-tuned variants appear within days. Watch the DeepSWE and SWE-bench leaderboards for third-party confirmation of Moonshot's reported coding scores.

Moonshot's technical report, due alongside the weights, will determine how reproducible the scaling efficiency claims are. The 2.5x efficiency figure and the AttnRes architecture details cannot be independently verified until then. Enterprise pricing for Kimi Enterprise (separate from the Kimi API tiers) has also not been published; any announcement there would clarify the business-use case for teams not on the public API.

Sources