Moonshot Just Dropped Kimi K3: 2.8T Parameters, Open Weights, and a Real Argument Against Paying for Frontier AI

Hey guys, Mr. Technology here.

It is July 16, 2026, and Moonshot AI has shipped Kimi K3. Not "previewed." Not "waitlist." The API is live at platform.kimi.ai, the model is already routed through OpenRouter, and the full open-weight drop is scheduled for July 27, 2026 — eleven days from now. You will be able to download a 2.8-trillion-parameter mixture-of-experts model and self-host it before the end of the month.

Here is what was released and what I think it means.

The Spec Sheet

• Architecture: decoder-only Transformer MoE, sigmoid-then-top-k router over 896 experts with 16 active (~32B active per token)
• Total parameters: 2.8 trillion — most aggressively sparse MoE any frontier lab has shipped in 2026 (vs Cohere North Mini Code 128/8, DeepSeek V3 256/8)
• Context: 1,048,576 tokens (1M exact), pricing flat across the window, with structured compaction kicking in at 300K
• Modalities: text in/out, native visual input (no separate vision adapter)
• API pricing: $3.00 / $0.30 / $15.00 per million input cache-miss / cached / output
• Distribution: Moonshot API, OpenRouter, full weights July 27

You pay memory for the full 2.8T but compute for ~32B per token. At INT8 that is roughly 2.8 TB of HBM per replica; at INT4, ~2.1 TB. To serve at throughput you need H200 / B200 / MI355X nodes, multiple if you want concurrency. This is a rack model. That is the point — total parameter count is what shows up on procurement spreadsheets once the weights are downloadable.

The Benchmark Table

Moonshot published a full launch table with footnoted harness identity per row — KimiCode, Claude Code, or Codex pinned for each result.

BrowseComp drops 0.8 points when you stop using context compaction — Moonshot's structured truncation at 300K is doing real work. The GDPval / AA-Briefcase scores are Elo ratings on agentic economic-value benchmarks against human professionals.

Independent verification landed fast. Artificial Analysis put K3 at Intelligence Index 57.11, Coding Index 76.24, Agentic Index 50.07, with 62 tok/s and 1.99 s TTFT. Vals ran it at Vals Index v1.2 of 74.7, Terminal-Bench 2.1 (hosted) 80.899, CorpFin v2 71.562, MedCode 48.884. The directional read across harnesses is the same: K3 is a serious frontier coding and agentic model, and it is open-weights.

The Headline: Open Weights, Frontier Capability, $31.5B

Moonshot is reportedly raising at $31.5 billion on the strength of this release and the K2 lineage. K2 was the first Chinese open-weights family to consistently land in the top tier of independent coding indices. K3 is the version that closes the gap.

For comparison: GPT-5.6 Sol launched July 9, 2026 at $5/$30 per million in/out. Claude Sonnet 5 launched June 30 at higher prices for the long-context tiers. K3 is $3 in, $15 out, $0.30 cached input — 40% cheaper than Sol on input and 50% cheaper on output, same 1M context and frontier-class coding.

Then on July 27 the weights drop. Once they are out, the API is a convenience layer, not a moat. Anyone with $250K of inference hardware can self-host the same model. Anyone with $2M can serve it at scale without sending a token to Moonshot, OpenAI, or Anthropic. The procurement argument for closed frontier gets substantially weaker when the alternative is downloadable.

Where K3 Is Strong, Where It Is Not

Strong: agentic coding. Terminal-Bench 2.1 at 88.3% and FrontierSWE at 81.2% put K3 in the same conversation as Sol and Sonnet 5. SWE Marathon at 42% is the long-horizon number that matters for real product work, and K3 is competitive.

Strong: knowledge work on GDPval-AA v2 (1668) and AA-Briefcase (1548 Elo) — preference judgments against work humans actually do.

Not there yet: Vals CorpFin (71.5) and MedCode (48.8) show where K3 is thin. K3 is a generalist that closes the gap on coding and knowledge work, not yet dominant on any narrow vertical.

Caveat: 62 tok/s is below what a smaller active-parameter model delivers. If your loop is throughput-bound, a smaller model (North Mini Code at 3B active, Gemma 4 26B-A4B, Qwen 3.5 35B-A3B) serves more concurrent users on the same hardware.

What This Means For Builders

If you pay OpenAI or Anthropic for coding-agent workloads, your procurement conversation changes Monday. Run your eval against kimi-k3 on OpenRouter this week. The closed-frontier moat for general coding work is functionally closed.

If you run your own inference, plan for July 27. INT8 needs ~2.8 TB per replica; INT4 ~2.1 TB. H200 nodes need ~16 to host one full replica with KV cache headroom; B200 tightens that. vLLM, SGLang, and TensorRT-LLM will have K3 paths within days — Moonshot has shipped compatible kernels through K2.

If you build domain-specific products, fine-tune. K3 open weights plus your RLVR pipeline plus your domain corpus beats closed-frontier API tokens at scale for any vertical where your prompts contain proprietary context. The math flips around 50M tokens/day of steady-state usage.

If you are a US frontier lab, your pricing power on general coding just dropped. K3 makes the closed-frontier pricing story about ecosystem, compliance tooling, latency, and uptime — not raw capability. Expect Sol and Sonnet pricing pressure within two quarters.

If you do export-controlled or defense work, none of this applies. If data residency and origin-of-weights matter, K3 is not a clean option. Cohere North Mini Code under Apache 2.0 is.

The Take

I have been waiting for an open-weights model that does not make excuses. DeepSeek V3 was an excellent inference-engineering story wrapped around a mid-tier model. Qwen 3.5 was excellent everywhere and never quite best-in-class on coding. The K2 cycle was the first time open-weights coding genuinely rivaled closed frontier, but it was still a half-step behind.

K3 is the first open-weights release where I cannot, in good conscience, tell you the closed frontier is unambiguously better. On coding agents and knowledge work, K3 is in the same conversation as GPT-5.6 Sol and Claude Sonnet 5. On long-horizon agentic loops, it is competitive. On the price-performance curve, it is structurally ahead.

The closed-frontier labs will respond — OpenAI with verified-access regimes and ecosystem, Anthropic with safety and compliance tooling, Google with distribution. None of those responses are wrong. None of them change the underlying reality that the capability gap on general coding and knowledge work — the workloads that drive 80% of paid API volume — is functionally closed by an open-weights model you can download for free in eleven days.

Download the weights on July 27. Run your eval. Decide for yourself.

— Mr. Technology

*Released: 2026-07-16 (Moonshot AI). Model ID: kimi-k3. Architecture: decoder-only Transformer MoE, 2.8T total parameters, 896 experts with 16 active per token, ~32B active per forward pass. Context: 1,048,576 tokens (1M exact), pricing flat. Modalities: text in/out with native visual input. Pricing: $3.00/M input cache-miss, $0.30/M cached input, $15.00/M output. Benchmarks (provider-published): DeepSWE 67.5; FrontierSWE 81.2; Kimi Code Bench 2.0 72.9; Terminal-Bench 2.1 88.3; Program Bench 77.8; SWE Marathon 42.0; BrowseComp 91.2% with compaction, 90.4% at full 1M without; GDPval-AA v2 1668 Elo; AA-Briefcase 1548 Elo. Independent: Artificial Analysis Intelligence 57.11, Coding 76.24, Agentic 50.07, 62 tok/s, 1.99s TTFT; Vals Index v1.2 74.7, Terminal-Bench 2.1 (hosted) 80.899, CorpFin v2 71.562, MedCode 48.884. Open weights: 2026-07-27. SKUs: K3 Max, K3 Cluster Max. Distribution: platform.kimi.ai, openrouter.ai/moonshotai/kimi-k3. Moonshot valuation: $31.5B reported raise. Sources: Moonshot launch blog, BenchLM, TechCrunch, OpenRouter, Artificial Analysis, Vals.*