Kimi K3: What Actually Matters in Moonshot's 2.8T-Parameter Model Beyond the Headline Number

Moonshot AI just released Kimi K3, and it might be the most important open-weight model of 2026. Forget the 2.8T parameters. That number is mostly marketing. Here is what actually matters: 1. Architecture: Kimi K3 is a Mixture-of-Experts (MoE) model with 2.8 trillion total parameters and a 1M-token context window. The two innovations worth understanding are: - Kimi Delta Attention (KDA) Instead of relying entirely on quadratic attention, KDA uses a hybrid linear attention mechanism that dramatically reduces compute and memory overhead on very long contexts. - Attention Residuals (AttnRes) Rather than propagating every hidden representation through all layers, AttnRes selectively reuses useful representations across network depth. 2. Where does it rank? Early independent rankings place it around #4 overall on Artificial Analysis. Current frontier looks roughly like: Claude Fable 5, GPT-5.6 Sol (High), GPT-5.6 Sol, Kimi K3 so it is ahead ahead of Claude Opus 4.8, and GPT-5.5 Moonshot's own evaluations also position K3 ahead of GPT-5.5 and Claude Opus 4.8 on several coding and agent benchmarks, though it still trails Claude Fable 5. As always, wait for more independent evaluations before treating benchmark wins as definitive. 3. Pricing This is where things get interesting (Input, Output) Kimi K3 - $3/M, $15/M GPT-5.6 Sol $5/M $30/M Claude Fable 5 $10/M $50/M For heavy coding agents or enterprise workflows, inference costs can become a major line item. K3 undercuts the flagship proprietary models while staying surprisingly competitive on quality. Try it out yourself on either of these: Kimi Chat, Moonshot API, OpenRouter, Kimi Code, Kimi Work Open-weight downloads are also expected, enabling self-hosting and fine-tuning for organizations that need data sovereignty. If the independent benchmarks hold up, Kimi K3 could become the default starting point for many production AI systems. #KimiK3 #OpenSourceAI #AgenticAI #AIEngineering

I checked this against Moonshot's official release materials and independent evaluations from Artificial Analysis.

Confirmed accurate:

• Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), two architectural updates designed to improve how information flows across sequence length and network depth, and it scales up MoE sparsity, activating 16 of 896 experts through a Stable LatentMoE framework — this matches the post's description of both innovations closely. Kimi
• The model contains 2.8 trillion total parameters but activates only a fraction of them per token due to sparse MoE routing, so the total parameter count doesn't describe the actual per-token compute — this directly supports the post's framing that "2.8T parameters... is mostly marketing" relative to what matters for real performance. Northflank
• Pricing is exact: Kimi K3 costs $3.00 per million input tokens and $15.00 per million output tokens, GPT-5.6 Sol costs $5 per million input tokens and $30 per million output tokens, and Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens — all three figures in the post match published rates precisely. Kie + 2
• The ranking claim checks out well: on Artificial Analysis' Intelligence Index v4.1, Kimi K3 scored 57.1, ranking fourth overall behind Claude Fable 5 (59.9) and GPT-5.6 Sol (58.9), but ahead of Claude Opus 4.8 (55.7), and its intelligence is comparable to Opus 4.8 and GPT-5.5 but remains behind Fable 5 and GPT-5.6 Sol. This matches the post's stated order almost exactly. IntuitionLabsX
• Moonshot did contribute a Kimi Delta Attention implementation to vLLM, supporting the self-hosting/fine-tuning framing, and open weights did follow. MarkTechPost

Worth flagging — mildly stale, not wrong:

• The post doesn't mention Claude Opus 5, which by early August 2026 leads the public Artificial Analysis Intelligence Index snapshot at 60.7%, ahead of Claude Fable 5 (59.9%). This isn't an error in the post so much as a sign it was likely written before Opus 5's benchmark entry, so the "current frontier" ranking is very slightly out of date rather than inaccurate at time of writing. BenchLM
• One nuance: the "open" model was API-only at launch, and weights weren't confirmed live on Hugging Face until closer to July 27, so "open-weight downloads are also expected" was accurate phrasing (expected, not yet available) rather than an overstatement. Eesel AI

Overall, this is an unusually well-grounded post — the architecture explanation, ranking, and all three pricing figures all check out against primary sources.