Kimi K3: Moonshot AI's 2.8 Trillion Parameter Open-Weight Model Redraws the Frontier
On July 16, 2026, Beijing-based Moonshot AI released Kimi K3 — a 2.8-trillion-parameter mixture-of-experts model that is now the largest open-weight AI model ever created, and the first Chinese model to break into the frontier pack on independent testing.
The numbers are staggering. With 2.8 trillion total parameters (activating 16 of 896 experts per token), a 1-million-token context window, native vision for images and video, and two entirely new architectural mechanisms — Kimi Delta Attention and Attention Residuals — K3 scores 57.1 on the Artificial Analysis Intelligence Index v4.1, placing it fourth overall behind only Claude Fable 5 (59.9) and GPT-5.6 Sol (58.9), and ahead of Claude Opus 4.8 (55.7).
This is not an incremental update. This is a reset of expectations for what open-weight models can achieve.
The Architecture Behind the Scale
Kimi K3 is a sparse mixture-of-experts model with 896 experts, of which only 16 are active for any given token. This sparsity is what makes a 2.8T parameter model computationally feasible — each token only passes through a fraction of the total parameters.
Two novel mechanisms power the architecture:
Kimi Delta Attention is a hybrid linear attention mechanism published by Moonshot on arXiv in late 2025. It combines the efficiency of linear attention with the quality of traditional softmax attention, achieving up to 6.3x faster decoding than prior approaches. The key insight is a delta-based formulation that approximates full attention while maintaining sub-quadratic complexity — critical for a model that needs to reason over a million tokens.
Attention Residuals, published separately in early 2026, serve as a drop-in replacement for standard residual connections. Where conventional residuals simply add the layer input back to the output, Attention Residuals learn a gated combination that preserves information more effectively through deep stacks. Moonshot reports consistent scaling gains from this technique across model sizes.
The model ships natively quantised at MXFP4 weights with MXFP8 activations, and Moonshot claims roughly 2.5x better scaling efficiency than the Kimi K2 generation — meaning K3 delivers more capability per unit of compute than any prior Moonshot model.
Benchmarks: Trading Blows with the Frontier
Independent testing broadly confirms Moonshot's claims. On the Artificial Analysis Intelligence Index, Kimi K3 debuts at 57.1 — the highest score ever recorded for a Chinese model, and higher than Claude Opus 4.8, Grok 4.5, GLM-5.2, and every open-weight model in existence.
On BrowseComp, a long-horizon information-seeking benchmark from OpenAI, K3 achieves 91.2 — a state-of-the-art score that speaks to the model's ability to navigate complex, multi-step research tasks across the web.
On GPQA Diamond, a graduate-level science reasoning benchmark, K3 scores 93.5%, matching GPT-5.6 Sol (94.1%) and besting Claude Fable 5 (92.6%). On Terminal-Bench 2.1, it hits 88.3% — level with GPT-5.6 Sol and ahead of Claude Fable 5's 84.6%.
On GDPval-AA v2, which measures economically valuable knowledge work across 44 occupations and 9 industries, K3 jumps to an Elo of 1,687 — up from Kimi K2.6's 1,190 and surpassing Claude Opus 4.8 (1,600), GLM-5.2 (1,514), and GPT-5.5 (1,494). It took the top spot on AutomationBench-AA (53%) and leads Arena.ai's Frontend Code arena ahead of Claude Fable 5.
Moonshot's own benchmark table (which should be treated as an upper bound) shows K3 winning about 7 of 35 published tests outright, and placing second or third in most of the rest, consistently behind Claude Fable 5 and GPT-5.6 Sol but ahead of Claude Opus 4.8 and GPT-5.5.
The Open-Weight Pivot: Strategic and Geopolitical
Kimi K3's release is as much a strategic move as a technical one. Full model weights are scheduled for release on July 27, 2026 — just 11 days after the hosted launch — under a licence that Moonshot has not yet published. If the weights ship as promised, K3 becomes the strongest open-weight model ever released by a wide margin.
The timing is no coincidence. Moonshot's strategic pivot to open-source began with Kimi K2 in July 2025 and accelerated through K2.5 and K2.6. The company was caught flat-footed by DeepSeek's rise in early 2025, seeing its monthly active user ranking in China slide from third to seventh. Open-sourcing is Moonshot's bid to recenter the global developer community around its ecosystem — a play that mirrors what Meta did with Llama and what DeepSeek did with V-series models.
For enterprise technology leaders, the implications are concrete: a frontier-capable model that can be self-hosted, fine-tuned, and integrated without API lock-in. The trade-off is infrastructure cost — running a 2.8T-parameter model requires substantial GPU capacity — but Moonshot's Mooncake project, which won Best Paper at FAST 2025, pioneered KV-cache-centric disaggregated serving designed specifically to make extreme-scale inference practical.
Pricing and Availability
Kimi K3 is priced at $3 per million input tokens, $15 per million output tokens, and $0.30 per million for cached input on the Kimi API and OpenRouter — roughly triple Kimi K2.6's pricing, but still below Claude Opus 4.8's $5/$25 and Fable 5's $10/$50.
Despite the sticker increase, Artificial Analysis puts K3's cost per task at $0.94 — about half Claude Opus 4.8's $1.80 and in line with GPT-5.6 Sol's $1.04 — because K3's efficiency means it completes tasks in fewer tokens.
In the consumer app, Kimi K3 is available on paid tiers starting at $19/month (Moderato, 256K context) with the full 1M window unlocked at $39/month (Allegretto). The model is accessible through kimi.com, mobile apps (iOS, Android, HarmonyOS), Kimi Work desktop agent, Kimi Code CLI, and the Kimi API with OpenAI SDK compatibility.
The Autonomous Agent Demo That Turned Heads
Beyond the benchmark numbers, Moonshot showcased a proof-of-concept that may be K3's most impressive achievement: designing a physical chip to run a nano-scale version of itself. Over 48 hours of continuous autonomous operation, K3 independently completed the full chip construction pipeline — architectural design, optimization, and verification — using open-source EDA tools. The result was a functional 4mm² chip design achieving timing convergence at 100 MHz and simulating over 8,700 tokens per second.
This is not a production chip. It is a demonstration of long-range autonomous capability — coherent, multi-step technical work sustained over two days. The same demo also included a computational astrophysics case where K3 reproduced the universal I-Love-Q relation, reading and cross-validating over 20 papers and implementing a complete numerical pipeline in approximately two hours — work that typically takes a senior researcher one to two weeks.
Caveats and Limitations
Kimi K3 is not without weaknesses. Independent testing by Artificial Analysis measured a hallucination rate of 51% — up on its predecessor even as accuracy improved. The model is verbose, generating roughly twice the output tokens of comparable models on equivalent tasks, which means real-world costs depend heavily on reasoning-token burn.
No SWE-bench Pro or SWE-bench Verified scores had been published at launch — Moonshot's coding claims rest on newer suites (DeepSWE, FrontierSWE, SWE Marathon) that are not yet independently replayable. And the open-weight release is a promise, not a shipped artefact: no checkpoint, licence, or model card exists as of this writing.
Moonshot itself acknowledges a UX gap versus Claude Fable 5 and GPT-5.6 Sol, sensitivity to preserved thinking history, and what it calls "excessive proactiveness" on ambiguous tasks — the model tends to take initiative even when caution would be warranted.
What It Means
Kimi K3 closes the performance gap between open-weight and proprietary models at the frontier. If the weights ship on July 27, every AI team in the world will have access to a model that competes with the best systems from companies with unfettered access to Nvidia's most advanced hardware — a development with implications that extend well beyond technology into geopolitics, enterprise strategy, and the structure of the AI industry itself.
The field now has four models at the very top: Claude Fable 5, GPT-5.6 Sol, Kimi K3, and Claude Opus 5. Three are proprietary. One is open-weight. That ratio may not hold for long.
Comments