Moonshot AI has launched Kimi K3, a 2.8-trillion-parameter model that the company calls “the world’s first open 3T-class model.” The model is available now on Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with full weights promised by July 27, 2026. It trails Claude Fable 5 and GPT 5.6 Sol on most benchmarks — but by margins that should terrify anyone betting the US frontier lead is safe.
🔍 THE BOTTOM LINE
Kimi K3 is not a catch-up model. It is a proof that the open-weight frontier, led by Chinese labs, has reached the parameter scale of the best proprietary systems — and is converting that scale into real intelligence at 2.5× the efficiency of its predecessor. The gap between open and closed is no longer a chasm. It is a crack.
A 2.8T Model You Can Download
The headline number is 2.8 trillion parameters, built on Moonshot’s Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) architecture. The model uses a Mixture of Experts (MoE) design that activates 16 of 896 experts per token — an extreme sparsity ratio that keeps inference costs competitive despite the total parameter count. According to Moonshot’s announcement, the architectural changes yield “approximately 2.5× improvement in overall scaling efficiency compared to Kimi K2.”
The model ships with native vision capabilities and a 1-million-token context window. Pricing on the Kimi API is aggressive: $0.30 per million tokens for cache-hit input, $3.00/MTok for cache-miss input, and $15.00/MTok for output. Moonshot claims a cache hit rate above 90% in coding workloads, which would make effective input costs closer to $0.30 than $3.00 for most real usage.
As the Financial Times reported, Kimi K3 “shows narrowing gap between US and China on frontier AI” — a framing that undersells what happened. This is not a narrow gap in one benchmark. It is a model that outperforms Claude Opus 4.8 and GPT 5.5 on multiple coding and agentic tasks, and trails Fable 5 by single-digit percentage points on most measures.
Where It Wins, Where It Doesn’t
The benchmark table tells a story that the FT headline doesn’t. On Kimi’s own evaluation suite, K3 scores are competitive across coding, agentic, reasoning, and vision:
- DeepSWE: 67.5 (K3) vs 70.0 (Fable 5) vs 73.0 (GPT 5.6 Sol) — third place, but ahead of Opus 4.8 (59.0)
- SWE Marathon: 42.0 (K3) vs 35.0 (Fable 5) vs 39.0 (GPT 5.6 Sol) — first place, a long-horizon coding test
- Terminal Bench 2.1: 88.3 (K3) vs 88.8 (GPT 5.6 Sol) — effectively tied
- BrowseComp: 91.2 (K3) vs 88.0 (Fable 5) vs 90.4 (GPT 5.6 Sol) — first place on web browsing
- GDPval-AA v2 (Elo): 1668 (K3) vs 1760 (Fable 5) vs 1748 (GPT 5.6 Sol) — third place, 5% behind
- GPQA-Diamond: 93.5 (K3) vs 92.6 (Fable 5) vs 94.1 (GPT 5.6 Sol) — effectively tied
The pattern is clear: K3 is not winning the agentic-Elo arms race, but it is winning or tying on long-horizon coding and browsing tasks. These are the tasks that map most directly to real engineering work. A model that wins SWE Marathon — a multi-hour, multi-file test — is not a toy.
The Agentic Claims Are Wild
Moonshot published case studies that would be extraordinary if independently verified. Kimi K3 reportedly:
- Optimized its own GPU kernel code, cutting forward+backward time from 283.6ms to 114.4ms across 15 hours of autonomous iteration
- Built “MiniTriton,” a compact Triton-like GPU compiler from scratch, including a tile-level IR, optimization passes, and PTX code generation — with performance “on par with or better than Triton” on roofline benchmarks
- Designed a chip in 48 hours to serve a nano model built on its own architecture, closing timing at 100 MHz in 4mm² with 1.46M standard cells
- Reproduced the I–Love–Q universal relations in computational astrophysics in “about two hours” — work that would typically take “one to two weeks”
These are claims from the company’s own blog, not independent evaluations. The kernel optimization result was compared against “Claude Fable 5 (with fallback),” which Moonshot flags as potentially including human-intervention steps. The chip design claim is a proof of concept, not a production chip. But the pattern — a model that can sustain long autonomous engineering sessions — is consistent with the SWE Marathon win.
The Architecture Bet
K3’s architecture is the real story for the AI research community. KDA and AttnRes are Moonshot’s contribution to the attention mechanism wars. The Stable LatentMoE framework, with Quantile Balancing and Per-Head Muon optimization, is what makes 896-expert training stable at trillion-parameter scale. The model uses MXFP4 weights with MXFP8 activations — quantization-aware training from the SFT stage onward — which means the shipped model is already optimized for inference, not just training.
Moonshot recommends deploying on “supernode configurations with 64 or more accelerators.” This is not a model you run on a Mac. But the open-weight release means any research lab or startup with adequate GPU infrastructure can run it — and fine-tune it, and study it, and build on it.
NZ Angle
New Zealand’s sovereign AI conversation has been about whether a small country can afford a seat at the frontier table. Kimi K3 changes the calculus. An open 2.8T model means the frontier is no longer locked behind an API paywall. A NZ research institution with access to a 64-GPU cluster — the kind of infrastructure being discussed in NZ sovereign AI planning — could run, study, and adapt this model. The open-weight frontier is an opportunity that did not exist six months ago.
The counterargument is the same one we have covered before: open weights from Chinese labs raise distillation and data sovereignty questions. Running Kimi K3 means trusting Moonshot’s training data provenance, which is not publicly documented. For research use, that is an acceptable risk. For production government use, it is not — yet.
The Open-Weight Frontier Timeline
K3 is the latest data point in a trend we have been tracking all year. For nine of the past twelve months, Kimi models have set the upper bound of open-model sizes. The trajectory:
- DeepSeek V4 Pro: 1.6T parameters
- Xiaomi MiMo V2.5 Pro: 1.02T
- Thinking Machines Inkling: 975B
- Z.AI GLM 5: 744B
- MiniMax M3: 428B
- Moonshot Kimi K3: 2.8T
The gap between the largest open model and the largest proprietary model is closing. Grok 5 at 6T parameters is still the biggest model we know about, but it is not open. K3 at 2.8T is open. The next open model to cross 3T will probably arrive before the end of 2026.
This is the trend that matters for the open-weights convergence thesis: open models are not just getting bigger, they are getting smarter per parameter. K3’s 2.5× scaling efficiency improvement over K2 means the next Kimi model could match Fable 5 without needing 6T parameters.
❓ FAQ
Can I run Kimi K3 locally? Not on consumer hardware. Moonshot recommends 64+ accelerator supernodes for inference. You can access it via the Kimi API ($0.30–$15/MTok), Kimi.com, or the Kimi Work desktop app. The full weights will be released by July 27, 2026, but running them requires serious GPU infrastructure.
How does Kimi K3 compare to DeepSeek V4 Pro? K3 is larger (2.8T vs 1.6T) and, based on Moonshot’s benchmarks, more capable on coding and agentic tasks. DeepSeek remains the open-weight coding champion for smaller-scale deployment; K3 is the open-weight frontier champion for labs that can afford the compute.
Is this a security risk? Open-weight models from Chinese labs are a subject of active policy debate. The concern is not the model itself — it is what the model can be fine-tuned to do. K3’s vision capabilities and 1M-token context make it useful for research, but also for disinformation at scale. The standard open-weight risk calculus applies.
Why does SWE Marathon matter more than other benchmarks? SWE Marathon tests multi-hour, multi-file software engineering — the closest benchmark to actual day-long developer work. K3 winning it (42.0 vs Fable 5’s 35.0) suggests it can sustain long engineering sessions without degrading. That is the skill that matters for AI coding agents in production.
🔍 THE BOTTOM LINE
The US-China AI gap is not a gap anymore. It is a race where the open-weight side, led by Chinese labs, is converting compute into intelligence more efficiently each generation. Kimi K3 is the model that makes the open-weights convergence thesis look less like speculation and more like a timeline. The weights drop July 27. The frontier is open.
