In partnership with
Your employees are connecting AI to everything. Now what?
ChatGPT and Claude don't just answer questions anymore. Employees are connecting them directly to Notion, Linear, Jira, and the rest of your stack. The AI can read, write, and take actions on company data. Most IT and security teams have no visibility into any of it.
Harmonic Security Connectors changes that. It sits inline with every AI-to-app connection, so you see each call, control what data moves, and block destructive actions before they happen. Employees notice nothing different.
See what's actually running across your business in a live demo.
Kimi K3: The 2.8T Open Monster That Built Its Own Chip
Moonshot AI just dropped the world's first open 3T-class model. It codes compilers, designs silicon, and reproduces astrophysics papers in 2 hours. Here's the full breakdown.
- 2.8 trillion parameters, 16 of 896 experts active, 1M token context. World's first open 3T-class model.
- New architecture: Kimi Delta Attention + Attention Residuals. 2.5x scaling efficiency over K2.
- Trails Claude Fable 5 and GPT 5.6 Sol overall, but beats Opus 4.8, GPT 5.5, and GLM-5.2 across coding and agentic benchmarks.
- Built MiniTriton (a GPU compiler from scratch), designed a physical chip in 48 hours, reproduced I-Love-Q astrophysics in 2 hours.
- API pricing: $0.30/MTok cache-hit, $3.00/MTok cache-miss, $15.00/MTok output. Weights drop July 27.
- Honest limitations: proactiveness issues, thinking-history sensitivity, UX gap vs Fable 5 and GPT 5.6 Sol.
Open Source Just Crossed the 3T Line
For nine of the past twelve months, Moonshot AI's Kimi models have set the upper bound of open-model sizes. Today, they smashed through the 3 trillion parameter ceiling with Kimi K3, a 2.8T-parameter Mixture-of-Experts model that activates only 16 of 896 experts at inference.
That is not a typo. 896 experts. 16 active. A sparsity ratio of 56:1.
But the raw parameter count is the least interesting thing about K3. The story is what this model does. It built a GPU compiler from scratch that rivals Triton. It designed a physical chip in 48 hours. It reproduced 1-2 weeks of astrophysics research in 2 hours. And it will be fully open-weighted by July 27, 2026.
Let's break down the architecture, the benchmarks, and the wild things this model pulled off.
K3 At a Glance
The Architecture: Delta Attention + Residuals
K3 introduces two architectural innovations that together deliver a 2.5x scaling efficiency improvement over K2. Here's the plain-English version.
1. Kimi Delta Attention (KDA)
Standard attention scales quadratically with sequence length. KDA changes the game by providing an efficient foundation for scaling attention across the 1M token context. The key innovation: it enables a new prefix caching approach that Moonshot has contributed back to the vLLM community, enabling competitive token pricing despite the massive model size.
2. Attention Residuals (AttnRes)
Instead of accumulating representations uniformly across depth, AttnRes selectively retrieves representations from earlier layers. Think of it as giving the model a "skip connection with memory" rather than a blind residual pass. The result: information flows more efficiently across model depth.
3. Stable LatentMoE + Quantile Balancing
With 896 experts and only 16 active, routing becomes a first-order challenge. Traditional MoE uses heuristic balancing losses with sensitive hyperparameters. K3 replaces this with Quantile Balancing, which derives expert allocation directly from router-score quantiles. No heuristics. No sensitive hyperparameters. Just math.
The Supporting Cast
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, rather than one-size-fits-all gradients.
SiTU (Sigmoid Tanh Unit): A new activation function that improves activation control over the standard SwiGLU or GELU approaches.
MXFP4 Weights + MXFP8 Activations: Quantization-aware training from the SFT stage onward. Broad hardware compatibility baked in, not bolted on.
Balanced Expert-Parallel Training: Static shapes, no host synchronization on the critical path. Prevents expert imbalance from killing throughput at large expert-parallel scales.
Moonshot recommends deploying K3 on supernode configurations with 64 or more accelerators. This is not a model you run on a single H100. The high-bandwidth communication domains are essential for inference efficiency at this expert count.
896 Experts, 16 Active: How The Routing Works
The Mixture-of-Experts design is the heart of K3's efficiency. With 2.8T total parameters but only 16 experts active per token, the model achieves massive capacity without proportional inference cost.
Benchmark Showdown: K3 vs The Frontier
Moonshot is refreshingly honest here. They straight up say K3 trails Claude Fable 5 and GPT 5.6 Sol overall. But the gap is narrower than you might think, and K3 beats everything else. Here's the picture across the key benchmarks.
Coding & Agentic Benchmarks
| Benchmark | Kimi K3 | Fable 5 | GPT 5.6 Sol | Opus 4.8 |
|---|---|---|---|---|
| DeepSWE | 67.3 | ~68 | ~65 | ~62 |
| Terminal-Bench 2.1 | Strong | Top | Strong | Mid |
| FrontierSWE | High | High | Top | Mid |
| PostTrain Bench | High | Top* | High | Mid |
| BrowseComp | 90.4 | ~89 | ~88 | ~85 |
*Fable 5 evaluated by third party, results may include fallback behavior. Scores are approximate where exact figures were not published in the blog. K3 uses KimiCode harness; others use their best-performing harness.
On BrowseComp, K3 scores 90.4 with a 1M token context window and no context management. That is a genuine frontier result. And in kernel optimization, K3 performed competitively with Fable 5 (with fallback) and substantially outperformed Opus 4.8, GPT 5.6 Sol, and GPT 5.5.
The Competitive Position
Relative positioning based on benchmark scores published in the Kimi K3 blog. Tier 1 = frontier proprietary, Tier 1.5 = near-frontier open, Tier 2 = competitive but behind.
4 Things K3 Did That Are Straight Up Insane
Benchmarks are one thing. But the demos Moonship showed are where K3 separates itself from every other open model. These are not toy tasks. These are real engineering work.
1. Built MiniTriton: A GPU Compiler From Scratch
K3 was asked to build a GPU programming system from scratch. It produced MiniTriton, a compact Triton-like compiler with its own tile-level IR layer over MLIR, optimization passes, and a PTX code-generation pipeline. Across roofline benchmarks, MiniTriton delivers performance on par with or better than Triton and torch.compile, beating Triton on certain workloads. It even sustains end-to-end nanoGPT training with stable convergence. A model built a compiler that rivals the compiler humans spent years optimizing.
2. 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. The result: a 4 mm² chip that closes timing at 100 MHz, sustains over 8,700 tokens/s decode throughput in simulation, packs 1.46M standard cells, 0.277 MB of SRAM, and an INT4 MAC array with fused dequantization. A chip designed by a model, for a model. This is long-horizon agentic capability at a level no open model has demonstrated before.
3. Reproduced I-Love-Q Astrophysics in 2 Hours
The I-Love-Q universal relations in computational astrophysics typically require 1-2 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 for exploring the results.
4. Edited Its Own Teaser Video 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. A high-density short video like this would typically take an experienced editor 1-2 working days, or a beginner 3-5. 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.
Why This Matters: The Open Frontier Is Real
The gap between open and proprietary is closing faster than anyone predicted. K3 is not quite Fable 5 or GPT 5.6 Sol, but it is close enough that the "open models can't compete" narrative is dead. When weights drop July 27, anyone with 64+ accelerators can run a near-frontier model.
3 Implications You Should Care About
1. The inference cost floor just dropped. At $0.30/MTok cache-hit with 90%+ cache hit rate in coding workloads, K3's effective pricing is absurdly low for a frontier-class model. This puts pressure on every proprietary API. If you are building agentic products, your unit economics just got significantly better.
2. Long-horizon agentic coding is table stakes now. K3 building a compiler, designing a chip, and reproducing astrophysics research is not a party trick. It is proof that the "agent that works for hours on hard problems" paradigm works. The question is no longer "can models do long-horizon work?" but "how do we orchestrate them safely?"
3. The 3T parameter mark is a psychological milestone. For 9 of the past 12 months, Kimi has set the open-model size record. K3 breaking 3T signals that open scaling is not slowing down. The proprietary labs no longer have a monopoly on the frontier.
- Kimi K3 Tech Blog
Pricing & Availability
K3 is live now on Kimi.com, Kimi Work, Kimi Code, and the Kimi API. Full weights drop July 27, 2026.
With a 90% cache hit rate, the effective input price is roughly $0.57/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. The Mooncake disaggregated inference architecture is doing serious work here.
Where to Get It
• Web: kimi.com (select Kimi K3)
• Desktop: Kimi Work v3.1.0+ (Windows, Apple Silicon Mac)
• Terminal: Kimi Code, select K3 via /model command
• API: Kimi API Platform, select kimi-k3
• Enterprise: Kimi Enterprise with data privacy and member management
• Weights: Full open weights by July 27, 2026
Limitations: What K3 Gets Wrong
Moonshot deserves credit for being upfront about K3's weaknesses. Too many model releases bury limitations in footnotes. Here are the real issues.
K3 was trained in "preserved thinking history mode." 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. Moonshot recommends using Kimi Code or another verified-compatible harness, and avoiding mid-session model switches.
K3's training emphasizes long-horizon, challenging tasks. The side effect: when it encounters minor issues or ambiguous user intent, it may make unexpected decisions on the user's behalf. If your application needs the agent to stay within well-defined boundaries, you need to impose explicit behavioral constraints in the system prompt or AGENTS.md.
Despite being highly competitive overall, K3 exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol. The model is capable, but the polish is not quite there yet. This matters for production deployments where UX quality directly impacts user retention.
The Verdict: A Genuine Frontier Open Model
Kimi K3 is not the best model in the world. Claude Fable 5 and GPT 5.6 Sol still hold that crown. But K3 is the best open model by a significant margin, and it is close enough to the frontier that the practical difference for most use cases is shrinking.
The real story is not the 2.8T parameter count. It is the agentic capability. Building a GPU compiler, designing a chip, reproducing astrophysics research, editing videos from raw clips. These are not benchmarks. They are real work. And an open model is doing them.
When weights drop on July 27, the game changes. The question is not whether open models can compete. They can. The question is what you build with them.
Next Issue Preview
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. Plus: what the K3 technical report reveals about the training data recipe.
Topic may shift based on developments. We cover what matters, not what we predicted.
