Sponsored by

One AI. Every Tool Your Store Actually Needs.

Most e-commerce sellers are paying for 6 to 8 separate tools that don't talk to each other — and spending hundreds of dollars a month just to keep up. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.

It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.

Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.

Free to start. No credit card required.

Start for free today

The Open Model That Nearly Beat Fable 5 Kimi K3 nearly beat Claude Fable 5. It is open. It built its own chip. Weights drop July 27.
researchaudio.io Issue 04 · July 17, 2026

The Open Model That Nearly Beat Fable 5

Kimi K3 is 2.8 trillion parameters, fully open, and trailing Claude Fable 5 by a razor-thin margin. It built a GPU compiler, designed a chip, and reproduced weeks of astrophysics in 2 hours. Fable 5 cannot do that. And the weights drop July 27.

Moonshot AI just crossed a line that every proprietary lab said would hold for another two years. Kimi K3 is a 2.8 trillion parameter open model. 896 experts. 16 active per token. 1 million token context. It trails Claude Fable 5 and GPT 5.6 Sol overall, but the gap is a sliver, not a canyon.

But the parameter count is the least interesting thing about K3. The interesting thing is what it did. It built a GPU compiler from scratch that rivals Triton. It designed a physical chip in 48 hours. It reproduced 1 to 2 weeks of astrophysics research in 2 hours. And Moonshot was honest about where it falls short.

Source: all data from kimi.com/blog/kimi-k3, published July 17, 2026.

 

The Numbers Nobody Is Publishing

Most coverage will say "2.8T parameters" and move on. The real story is in the ratios and the costs.

2.8T
Total Params
  
896 → 16
Experts: Total → Active
 
2.5x
Scaling Efficiency vs K2
  
$0.30
Per MTok (Cache Hit)
 

How It Actually Works

In plain English: Instead of one giant brain, the model has 896 smaller ones. Only 16 wake up per question. And instead of passing information forward blindly, each layer can reach back and grab exactly what it needs from earlier layers.

K3 introduces two architectural innovations that together deliver a 2.5x scaling efficiency improvement over K2. The first is Kimi Delta Attention (KDA), which provides an efficient foundation for scaling attention across the 1M token context. The second is Attention Residuals (AttnRes), which selectively retrieves representations from earlier layers rather than accumulating them uniformly.

The Mixture-of-Experts layer uses Stable LatentMoE with a new routing method called Quantile Balancing. Traditional MoE uses heuristic balancing losses with sensitive hyperparameters. K3 replaces this with math: expert allocation derived directly from router-score quantiles. No heuristics. No tuning. A shared expert is always active across all tokens.

FIG. 1 — K3 Block Architecture (simplified)

Each block contains KDA attention, Stable LatentMoE (16 of 896 experts), Attention Residuals, and Gated MLA. The dashed line is the AttnRes skip connection.

Four more innovations hold this together at 2.8T scale:

Per-Head Muon: Extends the Muon optimizer by optimizing attention heads independently. More adaptive learning at scale.

SiTU (Sigmoid Tanh Unit): A new activation function that improves activation control over SwiGLU or GELU.

MXFP4 weights + MXFP8 activations: Quantization-aware training from the SFT stage onward. Broad hardware compatibility baked in.

Balanced Expert-Parallel Training: Static shapes, no host synchronization on the critical path. Moonshot recommends deploying on supernode configurations with 64 or more accelerators.

 

The Scoreboard

Moonshot is refreshingly honest here. They straight up say K3 trails Claude Fable 5 and GPT 5.6 Sol overall. But the gap is narrow, and K3 beats everything else, including Opus 4.8, GPT 5.5, and GLM-5.2.

Benchmark K3 Fable 5 GPT 5.6 Sol Opus 4.8
DeepSWE67.3~68~65~62
Terminal-Bench 2.1StrongTopStrongMid
FrontierSWEHighHighTopMid
BrowseComp90.4~89~88~85
Kernel OptCompetitiveTop*BelowBelow

*Fable 5 evaluated by third party, results may include fallback. Scores approximate where exact figures not published. K3 uses KimiCode harness; others use their best-performing harness. Source: kimi.com/blog/kimi-k3.

The real flex

On BrowseComp, K3 scores 90.4 with a 1M token context window and no context management. On kernel optimization, K3 performed competitively with Fable 5 (with fallback) and substantially outperformed Opus 4.8, GPT 5.6 Sol, and GPT 5.5.

 

Four Things K3 Did That Are Straight Up Insane

Benchmarks are one thing. The demos are where K3 separates itself from every other open model. These are not toy tasks. These are real engineering work.

01 — GPU Compiler

Built MiniTriton from scratch

K3 built a compact Triton-like compiler with its own tile-level IR layer over MLIR, optimization passes, and a PTX codegen pipeline. Across roofline benchmarks, MiniTriton delivers performance on par with or better than Triton and torch.compile. It sustains end-to-end nanoGPT training with stable convergence. A model built a compiler that rivals the compiler humans spent years optimizing.

02 — Chip Design

Designed a physical chip in 48 hours

In a single 48-hour autonomous run, K3 built, optimized, and verified a chip using open-source EDA tools on the Nangate 45nm library. Within 4 mm squared, it closes timing at 100 MHz and sustains over 8,700 tokens per second decode throughput in simulation. 1.46M standard cells. 0.277 MB of SRAM. An INT4 MAC array with fused dequantization. A chip designed by a model, for a model.

03 — Astrophysics

Reproduced I-Love-Q in 2 hours

The I-Love-Q universal relations in computational astrophysics typically require 1 to 2 weeks of work by an experienced researcher. K3 did it in about 2 hours. It reviewed and cross-validated 20+ papers, implemented the full numerical pipeline, evaluated 300+ equations of state, identified inconsistencies in published formulas, generated 3,000+ lines of Python code, and produced an interactive HTML dashboard.

04 — Video Editing

Edited its own teaser from 56 clips

K3 edited its own teaser video from 56 source clips, handling clip selection, motion-matched cuts, frame-accurate beat synchronization, audio processing, and multiple rounds of revision. It also created a 3Blue1Brown-style motion-graphics explainer of its own architecture. The native multimodal architecture understands text, images, and video within the same model.

FIG. 2 — K3 vs Human Researcher: Time to Complete

The I-Love-Q reproduction is a 60 to 80x speedup over an experienced researcher.

 

Where It Works

• Long-horizon coding: Sustains engineering sessions, navigates massive repos, orchestrates terminal tools with minimal oversight.

• Vision-in-the-loop: Iterates between code and live screenshots for game dev, frontend, and CAD work.

• Research reproduction: Cross-validates 20+ papers, writes 3,000+ lines of code, produces interactive dashboards.

• BrowseComp: 90.4 with 1M context and no context management. A genuine frontier result.

Where It Collapses

• Thinking history sensitivity: If the agent harness fails to pass back all historical thinking content, or if you switch to K3 mid-session from another model, generation quality becomes highly unstable.

• Excessive proactiveness: When it encounters minor issues or ambiguous intent, it may make unexpected decisions on the user's behalf. You need explicit behavioral constraints in the system prompt or AGENTS.md.

• UX gap: Despite being highly competitive overall, K3 exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol.

 

The Pricing Math

K3 is live now on Kimi.com, Kimi Work, Kimi Code, and the Kimi API. Full weights drop July 27, 2026.

Cache-hit input$0.30/MTok
Cache-miss input$3.00/MTok
Output$15.00/MTok
Cache hit rate (coding)90%+

With a 90% cache hit rate, the effective input price is roughly $0.57 per MTok (90% x $0.30 + 10% x $3.00). For agentic coding workloads with heavy context reuse, this is genuinely cheap for a frontier-class model.

 

What This Means For You

For builders: The inference cost floor just dropped. At $0.57/MTok blended for a near-frontier model, agentic unit economics got significantly better. If you are building agentic products, your cost per task just shrank.

For engineers: Long-horizon agentic coding is table stakes now. The question is no longer "can models do long-horizon work?" but "how do we orchestrate them safely?" K3's proactiveness issue is the real production risk.

For the open ecosystem: The 3T parameter mark is a psychological milestone. For 9 of the past 12 months, Kimi has set the open-model size record. The proprietary labs no longer have a monopoly on the frontier.

The frontier is no longer behind a paywall. It is behind a download button.

 

Reader Challenge

1. When weights drop July 27, what is the first task you would hand K3?

2. Would you deploy K3 with its proactiveness issue, or wait for a constrained version?

3. At $0.57/MTok blended, what agentic workload becomes profitable that was not before?

 

Next Issue

When K3 weights drop July 27, we will have a full hands-on teardown: local deployment guide, inference cost analysis, and a side-by-side comparison with DeepSeek V4 and Llama 4 on the same hardware.

Topic may shift based on developments. We cover what matters, not what we predicted.

— researchaudio.io

Source: kimi.com/blog/kimi-k3 · July 17, 2026

Keep Reading