Moonshot AI announced Kimi K3 on July 16, 2026 — a 2.8 trillion parameter model that self-reported benchmarks place above Claude Opus 4.8 and GPT-5.5. The open-weight release planned for July 27 would make K3 the largest publicly available model in the history of Chinese AI labs.
Key takeaways
- 2.8 trillion parameters — more than 2x K2.6 and larger than DeepSeek v4 Pro (1.6T)
- benchmarks?benchmarks: standardized performance tests that measure model capabilities on defined tasks, enabling comparison across models Self-reported: outperforms Claude Opus 4.8 max and GPT-5.5 high; trails Claude Fable 5 and GPT-5.6 Sol
- Elo?Elo: a ranking system comparing models pairwise — a higher score indicates better relative performance against other models 1547 per Artificial Analysis; leads Arena.ai’s Frontend Code category
- API pricing: $3/1M input tokens, $15/1M output — up from $0.95/$4 in K2.6
- Moonshot valuation: $31.5B (up from $20B in May 2026)
The largest open-weight?open-weight: model whose weights are publicly available — can be downloaded, run locally, or fine-tuned without API restrictions model out of China
Kimi K3 packs 2.8 trillion parameters — more than twice K2.6 and ahead of DeepSeek v4 Pro (1.6T), the previous record holder among Chinese open-weight models. No official technical paper has been published yet, but the company confirmed vision input support and a 21% reduction in output tokens compared to K2.6 — a meaningful cost saving at this scale.
At launch, only the "max" inference?inference: the process of generating a response — unlike training, inference does not update model weights mode is available. Moonshot has not offered a variable-effort variant like those from Anthropic and OpenAI, which may limit cost flexibility for enterprise customers.
Benchmarks: strong, but not at the top
Moonshot's own data shows K3 ahead of Claude Opus 4.8 max and GPT-5.5 high on most tests. Independent platform Artificial Analysis puts it at Elo 1547 — a solid result. On Arena.ai, K3 leads the Frontend Code category, a strong signal for a model targeting developers.
The gap to the frontier is still visible: K3 trails Claude Fable 5 and GPT-5.6 Sol. Moonshot has not released granular per-benchmark breakdowns, making independent verification difficult — the numbers should be treated as claims until the open-weight release allows external evaluation.
Pricing: premium with open weights
K3 is priced at $3 per million input tokens and $15 per million output tokens — a steep jump from K2.6 and roughly equivalent to Claude Sonnet 5 pricing.
| Kimi K3 | Kimi K2.6 | |
|---|---|---|
| Input (USD/1M tokens) | 3.00 | 0.95 |
| Output (USD/1M tokens) | 15.00 | 4.00 |
A year ago, Chinese models were primarily positioned as cheaper alternatives to closed-source providers. K3 flips that narrative: Moonshot is targeting the premium segment, not price competition.
At the same time, the company is planning an open-weight release, meaning anyone can run K3 on their own infrastructure. The combination of premium API pricing and publicly available weights suggests Moonshot is targeting both enterprise managed API customers and teams looking to self-host on their own GPU cluster.
Chinese open-weight momentum
K3 lands in the context of a broader shift: Chinese open-weight models account for 41% of Hugging Face downloads and dominate the Top 6 on OpenRouter as of mid-July 2026. Moonshot, DeepSeek, and Meituan have all accelerated release cadence as enterprise demand for locally deployable models — free from third-party API data exposure — has grown.
Moonshot is simultaneously closing a new funding round at a $31.5B valuation, up 57% from its May 2026 round at $20B. K3 is both a technical product and an investor signal.
Why it matters
Kimi K3 is the first Chinese open-weight model to directly challenge Anthropic and OpenAI frontier models not just on price, but on benchmark performance. For the past two years, Chinese AI labs were largely seen as providers of cheaper alternatives — strong models, but clearly below Claude Opus or GPT-5. K3 blurs that line.
More consequential than the model itself may be the open-weight release. If a 2.8T model becomes publicly available, it shifts dynamics across the board: enterprise gets fine-tuning?fine-tuning: additional training on custom data to adapt a model for specific tasks without modifying weights from scratch access, the research community gets a model at this class, and Moonshot gets a reference position in the open-weight frontier segment.
The key question: will Moonshot's benchmarks hold up to independent verification once the weights are out? The July 27 release will answer that.
What's next
- Open-weight release of Kimi K3 scheduled for July 27, 2026 — enabling independent benchmarking and fine-tuning by external teams
- Moonshot finalizing a new funding round at $31.5B valuation — deal terms not yet disclosed
- No sub-max inference variant available yet — the company may address this before or after the open-weight launch
Sources
- TechCrunch — Moonshot's upcoming Kimi 3 is expected to close the gap with Anthropic's Opus 4.8
- Simon Willison’s Weblog — Kimi K3, and what we can still learn from the pelican benchmark
