Latent Space Jul 17, 2026 8/10 signal

[AINews] Kimi K3 2.8T-A50B: the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing

modelsindustry

Summary

Moonshot AI has launched Kimi K3, a 2.8 trillion parameter model positioned as the largest open-weights model ever. Independent evaluations place its performance in the class of Opus 4.8 and GPT-5.5, with a promised open-weights release by July 27, 2026.

Context

The release of Kimi K3 is a direct competitive move in the rapidly escalating race between open and closed-source AI models, particularly challenging recent attention given to models like Z.ai's GLM. It represents a significant attempt by the open-source community to close the performance gap with frontier closed models such as Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol. This launch is being framed as a milestone for open models, similar in impact to earlier influential releases from DeepSeek, by pushing the boundaries of scale and capability available to the public.

Details

Official Specifications and Features:

Model Size: 2.8 trillion total parameters.
Context Length: 1 million tokens.
Modality: Native multimodal input (text and images), with text-only output.
Architecture: Uses a LatentMoE / Stable LatentMoE architecture with 16 activated experts out of 896.
Key Technologies: Includes Kimi Delta Attention (KDA) for up to 6.3x faster decoding in million-token contexts, and Attention Residuals (AttnRes) for a claimed ~25% higher training efficiency at less than 2% additional cost.
Availability: Live on Kimi.com, Kimi Work, and Kimi Code, with an API. Open weights are promised by July 27, 2026.
Positioning: Marketed for long-horizon agentic coding, self-evolving workflows, and "vision in the loop" coding that iterates between code and screenshots.

Arena Benchmark Performance:

Arena	Metric	Kimi K3	Previous (K2.6)	Competitors
Frontend Code Arena	Rank	#1	#18	Surpassed Claude Fable 5
Frontend Code Arena	Score	1679 points	N/A	Ranked #1 in 6 of 7 domains
Frontend Code Arena	Pairwise Win Rate	76%	N/A	63% (Fable 5), 58% (GPT-5.6 Sol)
Text Arena	Rank	#9	#38	Top-10 in creative writing, coding
Text Arena	Score	1486 points	N/A	N/A

Artificial Analysis Independent Evaluation:

Benchmark	Kimi K3 Score	Comparison
AA Intelligence Index	57	Comparable to Opus 4.8 and GPT-5.5
GDPval v2	1668 Elo	N/A
AutomationBench-AA	53% (#1)	N/A
AA-Briefcase	1547 Elo	N/A
Cost per Task	$0.94	Used 21% fewer output tokens than K2.6

Acknowledged Limitations:

Moonshot AI officially stated that despite its competitive performance, Kimi K3 has a "noticeable gap in user experience" when compared to Claude Fable 5 and GPT-5.6 Sol.

What's new

The primary novelty is the sheer scale of the model: at 2.8 trillion parameters, Kimi K3 is claimed to be the largest open-weight model ever released, pending the delivery of the weights. It also introduces two named techniques, Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), aimed at improving decoding speed and training efficiency, respectively, at such a large scale.

Limitations

Moonshot AI explicitly acknowledges that Kimi K3 has a "noticeable gap in user experience" compared to top-tier closed models like Claude Fable 5 and GPT-5.6 Sol. Additionally, the 'open-weight' claim is currently a promise, with the weights scheduled for release on July 27, 2026, not available at the time of the announcement.

The take

Kimi K3's release is an aggressive move to commoditize frontier-level AI capabilities, directly challenging the performance and cost moats of closed-source leaders. By open-sourcing a model of this magnitude, Moonshot AI is forcing a conversation about whether massive-scale AI can be a public good rather than a private utility. The most telling detail is the self-admitted 'user experience gap' versus top competitors; this concedes that raw benchmark performance isn't the entire battle. Product polish, reliability, and the fine-tuning that creates a seamless user interaction remain key differentiators for closed models. Watch to see if the weights actually drop on July 27th and how quickly the open-source community can build on, fine-tune, and potentially close that UX gap.

Read the source →

Don't read this site daily. Get it in your inbox.

The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.