According to Inside, Moonshot AI (月之暗面) released Kimi K3, a 2.8-trillion-parameter open-weight model it says is the largest open-source release to date, which topped the Arena Frontend Code Arena leaderboard with 1,679 points. The launch coincided with a same-day 28.4% drop in Zhipu's Hong Kong-listed shares, and analyst Patrick Moorhead told CNBC the market's reaction carries a political dimension involving Washington's debate over Chinese open-weight models.
What are Kimi K3's core technical specifications, and why is it called a milestone for open-source models?
According to Inside, Kimi K3 has 2.8 trillion total parameters, which the outlet describes as setting a new record for open-source models. Inside also reports that K3 is 75% larger than its Chinese rival DeepSeek (深度求索) V4 Pro, which is estimated at roughly 1.6 trillion parameters.
Architecturally, Inside reports that K3 uses a Mixture-of-Experts (MoE) design containing 896 experts, of which only 16 are activated per token. The model also natively supports visual understanding and carries a 1-million-token context window, per Inside.
TechNews adds further architectural detail, reporting that K3 employs a new attention mechanism called Kimi Delta Attention (KDA), which mixes linear attention with a memory mechanism to reduce the burden of KV Cache. TechNews also reports that K3 pairs this with a technique called WideEP, which spreads the 896 experts across multiple GPUs so that each GPU's high-bandwidth memory (HBM) holds only a subset of experts — a design intended to optimize per-token memory usage and compute utilization.
How does K3 actually perform on international AI benchmarks?
According to Inside, K3 scored 1,679 points on the Arena's Frontend Code Arena public blind-test leaderboard, ranking first globally — ahead of Claude Fable 5 at 1,631 points and GPT-5.6 Sol at 1,618 points.
On a broader measure, Inside reports that K3 scored 57 on the Artificial Analysis composite intelligence index — which spans reasoning, coding, and knowledge tasks — placing third overall.
| Benchmark | Kimi K3 | Comparison | Source |
|---|
| Frontend Code Arena | 1,679 pts (1st) | Claude Fable 5: 1,631 pts; GPT-5.6 Sol: 1,618 pts | Inside |
| Artificial Analysis Index | 57 pts (3rd) | — | Inside |
What is K3's commercial pricing strategy, and what real-world cost issues has it raised?
According to Inside, K3's API pricing is set at $3 per million input tokens (dropping to $0.3 per million tokens on a cache hit) and $15 per million output tokens — which Inside describes as the highest price band any Chinese AI lab has ever set.
However, Inside also reports that independent testers found K3's reasoning-token consumption to be steep: generating a simple SVG graphic consumed more than 13,000 tokens, at a cost of roughly $0.25.
| Pricing Item | Value | Source |
|---|
| API input (standard) | $3 / million tokens | Inside |
| API input (cache hit) | $0.3 / million tokens | Inside |
| API output | $15 / million tokens | Inside |
| SVG generation (tester case) | ~13,000 tokens / $0.25 | Inside |
How did K3's launch immediately affect Chinese AI stocks and the broader industry?
According to Inside, the day K3 launched, Hong Kong-listed Chinese AI peers were hit hard: Zhipu (智譜) shares fell 28.4% and MiniMax shares fell 15.6%.
Inside also reports that Moonshot AI's own annualized recurring revenue has surpassed $200 million, and that the company is reportedly evaluating a Hong Kong IPO, with a new funding round said to target a valuation of up to $31.5 billion.
| Company/Metric | Figure | Source |
|---|
| Zhipu share price (day of launch) | -28.4% | Inside |
| MiniMax share price (day of launch) | -15.6% | Inside |
| Moonshot AI ARR | $200 million+ | Inside |
| Moonshot AI target valuation | up to $31.5 billion | Inside |
Why did K3's release draw political attention in Washington?
According to Inside, the U.S. House Foreign Affairs Committee held a hearing just two days before K3's launch to examine export-control gaps in what Inside describes as an "AI arms race."
Inside reports that analyst Patrick Moorhead told CNBC the market's dramatic reaction to K3 was partly political, noting that Washington is debating whether U.S. companies should use Chinese open-source models, and calling it "ironic" that "the Chinese seem to be doing just fine using their own models."
Inside also reports that Chinese President Xi Jinping, speaking at the World Artificial Intelligence Conference, called on nations to "encourage open-source and open collaboration, and cooperation and sharing," while opposing what he termed the "generalization of national security concepts" in AI.
On the technical-commentary side, TechNews reports that Dean Ball — a former Trump administration adviser now in a senior role at OpenAI — posted on X that K3 is "a very good model," adding: "I don't think its performance can be explained away by distillation or anything like that."
Yet TechNews also reports that Ball warned a world dominated by open-weight models could produce what he called "full AI communism," where AI is treated as a state-provided public good rather than a market product — which he described as aligning with China's stated position that AI should function as "a form of digital public infrastructure" provided by the state.
TechNews further reports that Ball predicted the Trump administration might eventually pursue a strategy of "manufacturing significant regulatory risk" around the use of Chinese open-weight models, reasoning that "you just need to create enough regulatory risk that every regulated enterprise backs off."
What are K3's practical application boundaries?
According to Inside, K3 has demonstrated a notable autonomous capability: in a 48-hour autonomous run, the model used open-source EDA tools to complete the design, optimization, and verification of a chip.
How does Moonshot AI plan to advance K3's open-source rollout and industry positioning?
According to Inside, Moonshot AI expects to release K3's full weights no later than July 27, with a technical report to follow. This rollout comes as Inside reports the company's annualized recurring revenue has already topped $200 million and it is reportedly weighing a Hong Kong listing alongside a funding round targeting a valuation as high as $31.5 billion.
What this means
The evidence lays out a striking juxtaction: Moonshot AI set the highest API price band ever charged by a Chinese AI lab (Inside, $3/$15 per million tokens) for a model it is simultaneously giving away as open weights (Inside, full release by July 27) — a model that immediately erased billions in market value from rivals Zhipu and MiniMax (Inside, -28.4% and -15.6%) while its own valuation ambitions climbed toward $31.5 billion (Inside). At the same time, a senior figure inside OpenAI's own orbit, Dean Ball, publicly praised K3's technical merit (TechNews) even as he warned that the same open-weight dynamic could reshape AI into a state-provided public good (TechNews) — a framing that appears to echo Xi Jinping's own call for "open-source and open collaboration" at the World AI Conference (Inside). That tension, layered onto a U.S. congressional hearing on export-control gaps held just two days before launch (Inside) and an analyst's on-record acknowledgment that market reactions carry political motives (Inside), suggests K3's story is being read simultaneously as a technical benchmark, a market shock, and a geopolitical signal.