According to VentureBeat, Moonshot AI has released Kimi K3, a 2.8-trillion-parameter model the company calls the largest open-source AI system to date. Full weights are set for July 27, 2026. On GDPval-AA v2 it scored 1,687, ranking third behind Claude Fable 5 Max and GPT-5.6 Sol Max, per the same report.
What are Kimi K3's basic specifications and technical characteristics?
According to VentureBeat, Moonshot AI released Kimi K3 as "a 2.8-trillion-parameter model that the company says is now the largest open-source AI model in the world" (E1). Full model weights are scheduled for release on July 27, 2026, based on technical documentation reviewed by researchers (E2). The model "features a 1-million-token context window, native visual understanding capabilities, and an always-on reasoning mode that the company calls 'thinking mode'" (E3).
Tom's Hardware reports that K3 uses a sparse mixture-of-experts design that "activates just 16 of its 896 experts per token, roughly 1.8% of the pool" (E16). Moonshot attributes efficiency gains to two architectural changes — Kimi Delta Attention, described as a hybrid linear attention scheme, and Attention Residuals, which alter how information moves between layers — claiming "roughly a 2.5x improvement in scaling efficiency over Kimi K2" as a result (E19).
How does Kimi K3 perform on major benchmarks, and how does it rank against rivals?
On GDPval-AA v2, VentureBeat reports that "Kimi K3 scored 1,687 — placing it third overall, behind only Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8), and ahead of Claude Opus 4.8 (1,600)" (E5). Separately, K3 "achieved a state-of-the-art score of 91.2 out of 100 on BrowseComp, a benchmark for long-horizon, high-difficulty information seeking" (E6).
Tom's Hardware adds that in the Arena Frontend Code evaluation, K3 was ranked "first ... at 1,679 points, ahead of Fable 5, in blind developer testing" on July 16, 2026 (E17). Moonshot itself describes K3 in its technical blog as "the world's first open 3T-class system and the largest open-weight AI model to date," while acknowledging that K3 "still sits behind Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol on overall performance" (E15).
| Benchmark | Kimi K3 | Comparison |
|---|
| GDPval-AA v2 | 1,687 (3rd) | Claude Fable 5 Max 1,815; GPT-5.6 Sol Max 1,747.8; Claude Opus 4.8 1,600 (E5) |
| BrowseComp | 91.2/100 | State-of-the-art score (E6) |
| Arena Frontend Code | 1,679 (1st) | Ahead of Fable 5 (E17) |
What changed in Kimi K3's API pricing compared to K2?
VentureBeat reports K3 "is priced at $3 per million input tokens and $15 per million output tokens, with cached input tokens dropping to just $0.30 per million" (E4). Tom's Hardware provides more granular figures: "$0.30 per million cache-hit input tokens, $3 per million on cache misses, and $15 per million output tokens," noting that "Kimi K2 launched a year ago at $0.60 per million input tokens, so uncached K3 input costs five times as much" (E18).
| Model | Input (cached) | Input (uncached) | Output |
|---|
| Kimi K2 (year-ago launch) | — | $0.60 / million | — |
| Kimi K3 | $0.30 / million | $3 / million | $15 / million |
What real-world capabilities has Kimi K3 demonstrated?
VentureBeat describes a 48-hour autonomous chip-design demonstration: "Over 48 hours of continuous autonomous agent operation, K3 independently completed the chip's full construction pipeline," producing "a tiny but functional chip design, just 4 square millimeters, that achieved timing convergence at 100 MHz and could decode more than 8,700 tokens per second in simulation" (E7). Tom's Hardware provides additional detail on the same case study, noting the design used "open-source EDA tools and the Nangate 45nm library," closed timing "at 100 MHz within 4mm squared," and "packed 1.46 million standard cells and an INT4 MAC array" while sustaining "more than 8,700 tokens per second of simulated decode" (E21).
On developer tooling, VentureBeat notes that "the Kimi Code CLI has accumulated over 3,100 stars on GitHub and features integration with VSCode, Cursor, and Zed" (E13).
How is Moonshot AI's fundraising and market position developing?
Moonshot AI was "founded in 2023 by Yang Zhilin, a Tsinghua University graduate who previously conducted research at Google and Meta," and quickly became a prominent Chinese AI startup (E8). VentureBeat reports that "by early 2026, it had raised roughly $1.5 billion across multiple rounds, with its valuation climbing from $2.5 billion to $4.3 billion and the company reportedly seeking a new round at $5 billion" (E9).
That growth followed a setback: "the release of DeepSeek's low-cost R1 model in January 2025 disrupted the entire Chinese AI landscape, and Moonshot AI was among the hardest hit. Kimi, which had ranked third in monthly active users in China, slid to seventh" (E10).
How large is Kimi K3's lead over open-source peers in scale?
According to VentureBeat, "the company's own timeline chart of open-source frontier model scale positions K3 as a dramatic outlier, towering above competitors like DeepSeek (1.6T), Xiaomi (1.02T), and Alibaba (397B)" (E11).
| Model | Parameters |
|---|
| Kimi K3 | 2.8T |
| DeepSeek | 1.6T |
| Xiaomi | 1.02T |
| Alibaba | 397B |
How has the release been received in China and internationally?
VentureBeat reports that "Xinhua, China's state news agency, framed the release as a national milestone, reporting that K3 'marks a new step forward in the development of China's artificial intelligence models'" (E14). Internationally, Tom's Hardware cites Bank of America analysts led by Alex Liu, who "said in a note cited by CNBC that K3 shows large-scale pre-training plus architectural work can still deliver step-change gains for flagship Chinese models despite compute constraints," in a note dated July 17, 2026 (E20).
What has Anthropic alleged about how Moonshot trained its models?
Tom's Hardware reports that "Anthropic accused Moonshot in February of using 3.4 million Claude exchanges to train its models through distillation, and K3 now benchmarks within a few points of the models named in that complaint" (E22). This allegation surfaced against a backdrop in which, as VentureBeat notes, "Anthropic disclosed in January [that] Claude Code reached $1 billion in annualized recurring revenue" (E12).
What does this add up to?
The evidence assembled across both reports shows a model that is simultaneously the largest by parameter count (2.8T, versus DeepSeek's 1.6T, Xiaomi's 1.02T, and Alibaba's 397B per E11) and, by Moonshot's own account in its technical blog, still third on overall performance behind Claude Fable 5 and GPT-5.6 Sol (E15), a ranking corroborated by the GDPval-AA v2 scores (E5). That gap sits alongside a pricing jump — uncached input costs five times K2's level a year earlier (E18) — and against the backdrop of Anthropic's distillation complaint, K3's benchmark proximity to the named models is notable given that the accusation itself dates to February, months before K3's July release (E22). Meanwhile, the scale achievement follows a period in which Moonshot's user rankings fell from third to seventh after DeepSeek's R1 release (E10), a decline that preceded the reported valuation climb from $2.5 billion to $4.3 billion (E9).