Kimi K3: Moonshot's 2.8T Open-Weight Model Sets a $3 Price Floor for the Frontier

Moonshot AI released Kimi K3 on July 16, 2026: a 2.8-trillion-parameter open-weight mixture-of-experts model with a native 1-million-token context window. It scores 88.3 on Terminal Bench 2.1, ahead of Claude Opus 4.8 at 84.6, and prices API input at $3.00 per million tokens. Full weights go public on July 27.

Frontier-class capability has lived behind closed APIs since the category emerged. Open-weight models earned a reputation as capable followers: they arrived quarters late and benchmarked a tier below the flagships from OpenAI, Anthropic and Google. Moonshot just compressed that gap to days and, on selected agentic benchmarks, erased it entirely. The commercial signal is equally loud: TechCrunch reports the company is raising fresh capital at a $31.5 billion valuation, up from $20 billion in May 2026, when it closed a $2 billion round. That is a 57 percent valuation jump in two months, powered by a model line investors now treat as frontier-grade. Moonshot built that credibility step by step: TechCrunch notes the Kimi K2 models already ranked near the top of open benchmarks, close behind the latest frontier releases. K3 converts closeness into parity on the tasks enterprises automate first. For enterprise buyers, the arithmetic of the vendor shortlist changed overnight.

What changed: 2.8 trillion parameters, priced to move

Kimi K3 is a mixture-of-experts system that routes 16 of 896 experts per token through Moonshot's Stable LatentMoE framework. The launch post details two architectural additions, Kimi Delta Attention and Attention Residuals, which together deliver roughly 2.5 times the scaling efficiency of Kimi K2. Weights ship in MXFP4 with MXFP8 activations, a quantization choice aimed squarely at affordable inference on current-generation accelerators. Scaling efficiency is the quiet number here: 2.5 times means Moonshot trains bigger models on the same compute budget, a structural advantage for a lab competing on cost. The 1-million-token context window is native rather than extended, which matters for contract analysis, codebase comprehension and long-horizon agent sessions.

The benchmark table is the headline. On Terminal Bench 2.1, K3 posts 88.3 against 84.6 for Claude Opus 4.8 and 84.6 for Claude Fable 5, landing within half a point of GPT-5.6 Sol's 88.8. On GPQA-Diamond it reaches 93.5, ahead of Fable 5's 92.6, and on MMMU-Pro it scores 81.6. US flagships keep clear leads elsewhere: GPT-5.6 Sol holds 73.0 on DeepSWE against K3's 67.5, and Fable 5 tops GDPval-AA v2 at 1760 versus K3's 1668. The pattern is legible: K3 wins terminal-style agentic coding, while closed frontier models defend complex software engineering and real-world occupational tasks.

Pricing lands harder than any single score. API input costs $3.00 per million tokens on cache miss and $0.30 on cache hit, with output at $15.00 per million. Cache-aware architectures get rewarded here: a 10x spread between cache-hit and cache-miss input pushes teams toward prompt designs that reuse context aggressively. The model is live today on Kimi.com, Kimi Work, Kimi Code and the Kimi API, and Moonshot has committed to publishing full open weights by July 27.

What it means for the vendor map: a published price floor for the frontier

An open-weight model matching closed flagships on agentic coding resets the reference price for the entire category. Every enterprise negotiation with Anthropic, OpenAI and Google now includes a credible alternative that costs a fraction per token and, from July 27, runs inside the buyer's own perimeter. Closed labs will defend their premium on agentic reliability, safety tooling, enterprise support and ecosystem depth — real differentiators that now get priced explicitly instead of bundled into the mystique of the frontier. Expect the familiar response playbook: aggressive cache pricing, batch discounts and enterprise bundles that shift the conversation from tokens to outcomes.

For hyperscalers and GPU clouds, K3 is pure demand: a frontier-class model that any AWS, Azure or CoreWeave customer can self-host expands the inference market far beyond proprietary API endpoints. For regulated European buyers, self-hosted weights answer the data-residency question directly, while provenance from a Chinese lab moves procurement review from legal formality to board-level conversation. TechCrunch reports enterprise executives already recommending open-source options from DeepSeek, Z.ai and Moonshot as alternatives to premium closed models. K3 upgrades that recommendation from cost-saving tactic to capability-neutral strategy on the benchmarks where it wins.

Watch the follow-through on July 27. Weight quality, licensing terms and the fine print on commercial use will determine whether K3 becomes enterprise infrastructure or stays an impressive API. The $31.5 billion valuation shows where investors have placed their bet.

The 90-day decision: run the bake-off before renewal season

Commission a two-workload bake-off and complete it before the end of Q3. Pick one long-context workload — contract review, codebase comprehension — and one agentic coding workload, then run Kimi K3 via API against your incumbent model. Measure cost per completed task, latency and human-escalation rate rather than cost per token: a cheap token that triples review effort becomes expensive. After the July 27 weights release, add a self-hosted deployment to the comparison and bring legal and compliance in on licensing and model provenance from day one. Buyers who arrive at their next vendor renewal holding K3 numbers negotiate from evidence, and the exercise creates leverage even when the incumbent wins. The frontier just became a market with a published price floor: $3.00 per million tokens, weights included.