Moonshot AI's Kimi K3 Beats Claude at Coding for 4.6x Less Money

Moonshot AI just dropped what may be the most consequential open-weights model release since DeepSeek R1. Kimi K3 is a 2.8-trillion-parameter sparse Mixture-of-Experts model that debuted at #3 on the DeepSWE leaderboard, scoring 69% on a benchmark designed specifically to stress-test frontier coding agents on long-horizon, real-world software engineering tasks. That puts it within a few percentage points of Claude Fable 5 (70%) and GPT-5.6 Sol (73%) -- the current top two -- and ahead of every other model on the list.

The benchmark that caught everyone's attention

DeepSWE, built by Datacurve, is a coding benchmark designed to avoid the saturation problem plaguing existing leaderboards. Today's leading public coding benchmarks are starting to saturate at the frontier, with top models clustering within a narrow score band. DeepSWE is built to separate them. The key design decisions that make it harder to game:

• Contamination-free: Tasks are written from scratch, not adapted from existing commits or PRs, so no model has seen the solution during pretraining.
• Real-world complexity: Prompts are roughly half the length of SWE-bench Pro's, yet solutions require 5.5x more code and ~2x more output tokens.
• Reliable verification: Verifiers are hand-written to test software behavior rather than implementation details.
• High diversity: Tasks span a broad pool of 91 repositories across 5 languages.

The benchmark also found serious problems with the status quo. SWE-Bench Pro, the current leading agentic coding benchmark, has a verifier that misgrades agent outputs at rates of 8% false positives and 24% false negatives. On DeepSWE, those rates drop to 0.3% and 1.1% respectively -- a meaningful improvement for anyone trying to make real decisions about which model to use.

2.8 trillion parameters, but only 16 are active at a time

Kimi K3 reports 2.8 trillion total parameters. It is a sparse Mixture-of-Experts model, routing each token through a small subset -- 16 of 896 routed experts -- rather than using all parameters densely for every token. Think of it like a massive panel of specialists where only a handful are consulted per question. This is what makes a model this large computationally tractable to serve.

The architecture introduces three new components worth understanding:

• Kimi Delta Attention (KDA): KDA is a hybrid linear attention mechanism. Moonshot states it enables up to 6.3x faster decoding in million-token contexts. Standard attention is quadratic in sequence length, meaning processing a million tokens is brutally expensive. KDA replaces that with a more efficient formulation that makes the 1M-token context window practically deployable rather than just theoretically possible.
• Attention Residuals (AttnRes): AttnRes selectively retrieves representations across depth rather than accumulating them uniformly. Moonshot states it delivers roughly 25% higher training efficiency at under 2% additional cost.
• Stable LatentMoE: At this level of sparsity, routing and optimization become first-order challenges. Quantile Balancing derives expert allocation directly from router-score quantiles, eliminating heuristic updates and a sensitive balancing hyperparameter. This is what keeps the model stable during training at this unprecedented scale.

Together with improvements in training methodology and data recipes, these structural advances give Kimi K3 roughly 2.5x the overall scaling efficiency of K2, converting compute into capability more effectively. Kimi K3 also applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility. Notably, KDA had a long incubation cycle: design reportedly started in January 2025 and took roughly 1.5 years to reach frontier scale.

Where it leads, where it trails

Kimi K3 tops LMArena's Frontend Code Arena at 1,679 points, a 17-place jump over Kimi K2.6, but lands a more modest #9 on the general Text Arena. The coding story is strong; the general-purpose story is more nuanced.

On DeepSWE specifically, the numbers tell a clear story about cost efficiency. Kimi K3 achieves 69% at an average cost of $4.65 per task, compared to Claude Fable 5's 70% at $21.63 per task. That's roughly a 4.6x cost advantage for near-identical performance on this benchmark. Moonshot's own internal tests acknowledge that K3's performance still trails GPT-5.6 Sol and Claude Fable 5 in some areas, though the company says it is extremely close behind those models in several key tasks.

The honest caveat: early evaluations place Kimi K3 second in performance metrics when benchmarked against top-tier western models, but independent benchmarks beyond provider-reported scores remain largely absent, so those rankings deserve a healthy grain of salt until third-party testing catches up.

Pricing and access

The launch surfaces span the consumer product (Kimi.com), the workplace tier (Kimi Work), the terminal coding agent (Kimi Code), and the developer API -- all live on day one. You can try it immediately at kimi.com without a credit card.

API pricing is structured as follows:

• $0.30 per million cache-hit input tokens, $3 per million uncached input tokens, and $15 per million output tokens.
• For comparison, Fable 5 costs $1 per million input tokens and $50 per million output tokens, while GPT-5.6 Sol costs $0.50 per million input tokens and $30 per million output tokens.

To call the API today:

from openai import OpenAI
client = OpenAI(
    api_key="YOUR_KIMI_API_KEY",
    base_url="https://api.moonshot.ai/v1"
)
response = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Refactor this Go function to handle edge cases..."}]
)
print(response.choices[0].message.content)

K3 is available in Kimi Code from the Moderato plan up, at 256K context; the full 1M window requires Allegretto or above.

The open-weights catch

At 2.8 trillion parameters, K3 is being called the largest open-weight model shipped to date. The company's own timeline chart of open-source frontier model scale positions K3 as a dramatic outlier, towering above competitors like DeepSeek (1.6T), Xiaomi (1.02T), and Alibaba (397B). But there's an important nuance: Kimi K3 is open-weight, but you cannot download it yet. Moonshot has committed to releasing the full weights by July 27, 2026 on Hugging Face; until then you reach K3 through the app and API only.

Self-hosting is not a realistic option for most teams regardless. At 2.8T parameters -- even in MXFP4 -- Moonshot's own serving recommendation is a supernode of 64 or more accelerators. This is not a self-host-on-your-own-GPU story. For nearly every organization, "running K3 yourself" will mean renting managed capacity from an inference provider that hosts the open weights -- the same consumption model DeepSeek V4 established.

Why this matters beyond the benchmark

Kimi K3's release forces a recalibration of several assumptions that have guided enterprise AI strategy. The performance gap between open-source and proprietary models has functionally closed at the frontier. That has real consequences for pricing leverage, vendor lock-in, and the self-hosting roadmap for any team building on top of closed APIs.

China's AI ecosystem, which many Western observers questioned after early struggles with chip export restrictions, has now produced a model that competes with the best systems from companies with direct access to Nvidia's most advanced hardware. The architectural innovations behind K3 -- particularly the hybrid linear attention mechanism -- suggest that algorithmic efficiency may matter as much as raw compute.

The rollout was notably informal -- described by one observer as "no keynote, no model card, just a quiet overnight flip of kimi.com" -- echoing the low-key release strategy that DeepSeek used to introduce R1. The reaction was anything but quiet: rival AI stocks fell sharply, and former White House AI policy adviser Sriram Krishnan called it "a big moment, with multiple implications for the entire industry."

The July 27 weights drop will be the real test. If independent evaluations confirm the benchmark numbers, K3 becomes the most capable freely available model in the world -- and the argument for paying frontier closed-model prices gets a lot harder to make.