Summary
- On July 16, Moonshot AI unveiled Kimi K3, a massive 2.8-trillion-parameter open-weight model that surpasses U.S. counterparts in certain specialized benchmarks.
- Priced similarly to Claude Sonnet 5 at $3 per million input tokens and $15 per million output tokens, K3 achieves scores closer to Fable 5.
- The complete model weights will be available by July 27 under a modified MIT license, making K3 the largest open-source AI model ever.
Moonshot AI has released what is now the most significant Chinese open-source model, Kimi K3, which outperformed Claude Fable 5 in script-writing tasks.
The Towards AI's Writing Elo benchmark, which evaluates models based on their ability to write scripts that are then judged against published works, ranked Kimi K3 at 2,840, surpassing Fable 5's score of 2,760. Historically, Anthropic's team has dominated this ranking.
Exciting news from our internal writing benchmark (initial results): Kimi K3 by @Kimi_Moonshot is now the top performer in our editorial voice, achieving an Elo score of 2840, outpacing Claude Fable 5.
This marks a rise from #21 to #1 compared to its predecessor, Kimi K2.6, with a cost of about $0.25 per script: five times more efficient… https://t.co/000EosInEc pic.twitter.com/V18t7g4pHu
— Louis-François Bouchard 🎥🤖 (@Whats_AI) July 16, 2026
K3 also topped Arena AI's Frontend Code Leaderboard, a ranking determined by thousands of human votes on code generation tasks, with a score of 1,679 compared to Fable 5's 1,631. It ranked first in six out of seven frontend categories.
Breaking news: Kimi K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 points, outpacing Claude Fable 5.
This represents a 17-place improvement from Kimi K2.6, moving from #18 to #1.
In Frontend, Kimi K3 secured the top position in six out of seven domains: Brand & Marketing, Reference-Based Design, Data & Analytics,… https://t.co/YDN3BufGkC pic.twitter.com/Oa6teaQnWp
— Arena.ai (@arena) July 16, 2026
The Artificial Analysis Intelligence Index, which aggregates scores from nine independent assessments covering coding, reasoning, agentic tasks, and knowledge, rates K3 at 57. For comparison, Claude Fable 5 is at 60, GPT-5.6 Sol at 59, and Claude Opus 4.8 at 56, placing K3 as the third most capable model overall, trailing Fable 5 by just 3%.
Kimi K3 is outperforming Fable 5 in direct comparisons on BridgeBench.
In the same task, judged by a blind panel, K3 triumphed in 7 out of 8 categories, winning Refactoring 9-0 and Debugging 6-1.
Fable 5's sole victory was in Speed.
Fable 5 scored 142 to 36 against GPT 5.6 Sol but is now trailing 2 to 1 to an open-source model… pic.twitter.com/CaRSjbIwUW
— Bridgebench (@bridgebench) July 17, 2026
Kimi K3 is leading in performance on https://t.co/SnZ54Xok7n, surpassing Fable and achieving a similar success rate in less time.
This marks the first occasion an open model has outperformed all proprietary ones in this comprehensive web engineering benchmark.
Notes:
▪️ Benchmarks… pic.twitter.com/2MqvbAxAaw
— Guillermo Rauch (@rauchg) July 16, 2026
For an example of K3's capabilities, a zero-shot result prompt requested the model to create an iOS clone. In comparison, this is the best approximation shared on social media using GPT 5.6 Sol with a more complex prompt.
Understanding K3's Design
K3 features 2.8 trillion parameters—numerical values that encapsulate a model's knowledge—utilizing a mixture-of-experts architecture. This approach divides parameters into 896 "expert" subnetworks, activating only a portion for specific tasks, allowing for advanced intelligence without overwhelming server capabilities.
According to Moonshot AI, "It is the world’s first open-source model in the 3-trillion-parameter class, designed for frontier intelligence scenarios including long-horizon coding, knowledge work, and reasoning." This claim is backed by evidence: DeepSeek's V4-Pro has a maximum of 1.6 trillion parameters, while Moonshot's own K2 is at one trillion. K3 nearly doubles the size of its nearest open-weight competitor.
It offers a one-million-token context window—tokens being the fundamental units of information processed by AI, roughly equivalent to three-quarters of a word—along with native capabilities for understanding images and videos, plus continuous reasoning.
Efficiency improvements are supported by two architectural innovations. Kimi Delta Attention enhances decoding speeds for long sequences, achieving up to 6.3 times faster processing at million-token contexts. Attention Residuals allow for selective information routing across model layers, rather than uniform accumulation, increasing training efficiency by about 25% at under 2% extra computational cost—resulting in roughly 2.5 times better scaling efficiency compared to K2.
Pricing and Benchmark Performance
Kimi K3 is priced at $3 per million input tokens and $15 per million output tokens, matching the pricing of Claude Sonnet 5, Anthropic's mid-range model. However, while Sonnet 5 serves as Anthropic's average offering, K3 sits three points below Fable 5 on the Artificial Analysis Index. When assessed across the nine-benchmark suite, K3 costs $0.94 per task, compared to $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8.
Essentially, this model delivers premium performance at mid-tier pricing.
As reported by Decrypt in May, the price difference between Chinese and American frontier AI was previously 15-30 times. K3 does not undercut DeepSeek's rates but is priced like a Western mid-range model, offering nearly frontier-level performance within that bracket. This represents a significant cost advantage for teams utilizing the API.
If Anthropic proceeds with plans to limit Fable 5's availability to API access only, K3 will become the closest open-weight alternative to the leading industry model, offered at half the per-task cost of Opus 4.8. This scenario is already being calculated by benchmark enthusiasts.
The launch of K3 raises questions for advocates of U.S. chip export controls. The U.S. restricted Nvidia's H800 GPUs from being exported to China in late 2023; Moonshot acknowledged that earlier models were trained using those chips. K3's benchmark documentation references H200s and what is described as "a GPGPU from an alternative vendor," widely thought to be Huawei Ascend hardware, though the specifics remain unclear.
Yutong Zhang, president of Moonshot AI, addressed the constraints directly at Davos this year, stating as per Silicon Republic: "We knew we didn't have the luxury to simply scale up compute… That forced us to focus on fundamental research and efficiency." Analysts at Bank of America noted post-launch that K3 illustrates how "pre-training scaling, combined with architectural innovation, can still yield significant advancements for flagship Chinese models" despite these limitations.
Moonshot is part of the so-called AI Tiger startups, which have collectively transformed the global model landscape without the chips that the U.S. claimed were essential. Whether this serves as a case for stricter export controls or demonstrates their ineffectiveness remains a question for policymakers in Washington.
Critical Considerations
K3's hallucination rate on the AA-Omniscience benchmark—a measure of how often a model confidently produces incorrect answers—rose from 39% to 51% compared to its predecessor K2.6. While it provides more accurate responses overall, it also generates more inaccuracies. The model's documentation acknowledges that it can be "excessively proactive," potentially making unexpected choices on behalf of users during lengthy autonomous tasks.
For teams that utilized the Kimi K2.6-based tools and are considering an upgrade, K3 offers significant improvements in many areas, but the increase in hallucination rate should be carefully evaluated before trusting it with tasks requiring high accuracy.
If you're interested in trying K3 for free, it is accessible on Kimi’s official website. However, heavy server traffic may lead to frequent task interruptions, making it difficult to use. A more reliable option is to pay for a subscription or access it via API.
Model weights are set to be released on July 27, which will be available for large enterprises and organizations. Currently, no domestic GPU can handle a model of this magnitude.