Kimi K3 and Inkling: The Week Open Weights Reached the Frontier — and What It Takes to Run Them

In the space of about thirty hours in mid-July 2026, the open-weight frontier stopped being a runner-up category. On July 15, Mira Murati's Thinking Machines Lab shipped Inkling, its first in-house model and the largest American open-weight release to date. The next day, Beijing's Moonshot AI answered with Kimi K3 — a 2.8-trillion-parameter behemoth the company bills as the first "open 3T-class" system and the largest open-weight model ever announced. Markets reacted the way they did to DeepSeek a year and a half earlier: Fortune and Axios both framed K3 as a fresh shock to the assumption that U.S. labs stay ahead simply by outspending everyone on compute.

Both models are genuinely powerful. Both are also enormous, and the more interesting story for anyone who actually wants to use them is the second one: the size of the machine you need under the model. This post breaks down what Kimi K3 is, what it takes to run it well, how Inkling compares, and where the two sit inside a July that also delivered GPT-5.6, Sonnet 5, and a wave of agentic products.

A note on sourcing. Kimi K3's weights are not public yet — Moonshot has promised them by July 27. Until then, every benchmark number below is a Moonshot-reported figure, and every memory/GPU footprint for K3 is a community estimate scaled from the K2 family. We flag which is which throughout. Inkling's numbers, by contrast, are already verifiable against published weights.

Kimi K3: the power

Start with the architecture, because it explains both the capability and the cost.

K3 is a Mixture-of-Experts model with 2.8 trillion total parameters but only 16 of its 896 experts active per token — roughly 1.8% of the pool on any given forward pass, per Moonshot's launch write-up as reported by Tom's Hardware. That extreme sparsity is the whole trick: you get the knowledge capacity of a 2.8T model with the per-token compute of something far smaller.

Two architectural changes do the heavy lifting, and Moonshot claims they deliver roughly a 2.5× improvement in scaling efficiency over Kimi K2:

• Kimi Delta Attention (KDA) — a hybrid linear-attention scheme that cuts the quadratic cost of long context.
• Attention Residuals — a change to how information moves between layers.

Moonshot also ran quantization-aware training from the supervised fine-tuning stage onward, using MXFP4 weights and MXFP8 activations — a combination it says it chose specifically for broad hardware compatibility. That detail matters later. The model ships a 1,048,576-token context window (an exact power-of-two megatoken) with native vision, and Moonshot says pricing stays flat across the whole window rather than stepping up through context tiers.

The benchmarks (Moonshot's own)

Moonshot is refreshingly candid that K3 does not top the closed frontier: it says K3 sits behind Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol on overall performance. But it claims K3 beats every other model in its evaluation suite — including Claude Opus 4.8 and GPT-5.5 — across coding and agentic work. The headline results, as compiled by VentureBeat:

Additional coding figures from Moonshot's table: DeepSWE 67.5, FrontierSWE 81.2, and an internal Kimi Code Bench 2.0 of 72.9. Independent watcher Simon Willison put K3 through his informal "pelican on a bicycle" test the day it dropped — a reminder that until the weights land, everyone is testing a hosted endpoint, not the model.

Two caveats belong right next to those numbers. First, as Tom's Hardware notes, every K3 figure is currently a Moonshot claim — reported or drawn from API access — and can't be independently reproduced until July 27. Second, there is an unresolved controversy: Anthropic accused Moonshot in February of distilling roughly 3.4 million Claude exchanges to train its models, and K3 now benchmarks within a few points of the very models named in that complaint. Treat the leaderboard placements as marketing until the community reproduces them — the same discipline we applied to June's open-weight surge.

Pricing

K3's API runs $0.30 per million cache-hit input tokens, $3 per million on cache misses, and $15 per million output tokens. That's a striking move: it puts K3 at roughly Claude Sonnet tier and makes it the most expensive model any Chinese lab has shipped, a big jump over Kimi K2.6's $0.95/$4. Moonshot is betting that "open and near-frontier" is worth premium pricing — a bet that only holds if the self-hosted alternative is expensive too. Which brings us to the hardware.

Kimi K3: the hardware reality

This is the part most launch coverage glosses over, and it's the part that decides whether K3 is relevant to you.

Moonshot's own deployment guidance is unambiguous: this is not a single-rack model. Per the launch write-up relayed by VentureBeat and Tom's Hardware, Moonshot recommends serving K3 on supernodes of 64 or more accelerators, keeping expert-parallel traffic inside a single high-bandwidth domain. With 896 experts sprayed across the network, the moment your expert-parallel routing has to cross a slow interconnect, throughput collapses. A 2.8-trillion-parameter MoE wants a lot of fast-networked GPUs sitting physically close together.

For memory, we have to estimate — nobody outside Moonshot and its launch partners has the weights. Scaling from Unsloth's published build sizes for the 1T K2 family (the closest reference, since K2 ships its MoE weights in native INT4), the modemguides hardware analysis lands here:

Read that bottom row carefully: even the most aggressive quantization of K3 is likely to need 650 GB or more, which is past every consumer ceiling there is — including a maxed-out 512 GB Mac Studio. In practical terms, "run K3 well" means multiple 8-GPU H200- or B200-class nodes — realistically 16+ high-end accelerators — not a workstation. The modemguides piece puts it bluntly: K3 is "not a home setup. It is a rack line item." (Moonshot has not disclosed K3's active-parameter count; press estimates range from 40B to 100B, which affects speed but not the memory wall.)

One hardware nuance worth understanding if you're budgeting: Moonshot chose MXFP4/MXFP8 for its quantization-aware training precisely because it wants K3 to run across a broad hardware base. But the most memory-efficient FP4 formats (NVFP4) require NVIDIA Blackwell silicon — compute capability 10.0 and above, i.e. B200, GB200, or B300. Hopper-generation H200s (compute 9.0) lack the FP4 tensor-core hardware entirely. So the cheapest dense path to serving a model this size increasingly assumes you've bought into Blackwell, not last generation's H100/H200 fleet.

Bottom line on K3: it is arguably the most capable open-weight model in existence, and it is priced — in both API dollars and self-hosting iron — like a frontier model rather than a hobbyist one. The "open" in "open weights" here means auditable and ownable, not cheap to run.

Inkling: America's answer, with real numbers you can check today

Twenty-four hours before K3, Thinking Machines Lab — the startup founded by former OpenAI CTO Mira Murati — shipped Inkling, and did the thing Moonshot hasn't yet: released the weights immediately, under Apache 2.0, on Hugging Face.

The specs:

• 975 billion total parameters, 41 billion active — a Mixture-of-Experts transformer.
• 1-million-token context window.
• Natively multimodal: pretrained from scratch on 45 trillion tokens of text, images, audio, and video, with capabilities spanning reasoning, coding, tool use, visual analysis, speech transcription, and long-form audio understanding.

Because the weights are public, the hardware numbers are confirmed, not estimated — and this is the cleanest side-by-side you'll get for what a ~1T open model actually costs to serve. Per The Register and the Databricks deployment notes:

That NVFP4 checkpoint — shipped alongside the original, optimized for Blackwell — is the story for most teams: it halves the GPU count. And notice the pattern repeating from K3: FP4 efficiency is a Blackwell privilege.

Where Inkling diverges from Moonshot is philosophy. Thinking Machines is explicit that Inkling is not the strongest model, open or closed. Its pitch is that it's the best base to build on:

• Calibrated by design — it flags uncertainty rather than guessing, and lets you dial "thinking effort" up or down to trade quality for speed.
• Token-efficient — on one coding benchmark the company says Inkling uses a third as many tokens as NVIDIA's Nemotron 3 Ultra to reach the same score, which is the kind of efficiency that shows up directly on your inference bill.
• Censorship-resistant — Thinking Machines trained it to answer directly on politically sensitive topics; Cognition's Propaganda and Censorship Eval flagged strong non-compliance with censorship patterns, a deliberate contrast with the Chinese open models.
• Built for fine-tuning — customization runs through the lab's Tinker platform, with inference supported on SGLang, vLLM, and llama.cpp, plus Together AI, Fireworks, Modal, Databricks, and Baseten.

On raw benchmarks, VentureBeat has it landing between Kimi 2.5 and 2.6 overall — ahead of NVIDIA's Nemotron 3 Ultra (Inkling posts 97.1% on AIME 2026 and 77.6% on SWEBench Verified vs Nemotron's 94.2% and 70.7%, and 74.1% on MCP Atlas vs 44.7%), but trailing China's GLM 5.2 on pure coding. The training run was underwritten by a multi-year NVIDIA partnership: roughly 1 gigawatt of compute and access to next-generation Vera Rubin systems.

Inkling-Small: the one to watch

Alongside the flagship, Thinking Machines previewed Inkling-Small — 276 billion total parameters, 12 billion active — and buried the most interesting line in the announcement: it matches or exceeds the full 975B Inkling on many benchmarks, thanks to an improved pre-training recipe. At 12B active params it targets latency-sensitive workloads, and its full weights follow after testing. If a 276B model really does trade blows with a 975B one, the practical open-weight frontier just got a lot cheaper to serve — no supernode required.

The rest of the July 2026 wave

K3 and Inkling didn't land in a vacuum. July was one of the densest model months on record, per the ThursdAI and llm-stats trackers:

• OpenAI — GPT-5.6 (Sol / Terra / Luna). A three-tier lineup rather than one model: Sol for high-end reasoning and science at $5/$30 per million, Terra at GPT-5.5 quality for half the cost, Luna for cheap high-volume work. All ship a 1M-token context and a February 2026 cutoff. Sol became the first model to beat a public game on ARC-AGI-3 (7.8%) — though METR flagged the highest benchmark-cheating rate it has ever measured during pre-deployment testing.
• OpenAI — GPT-Live (full-duplex voice). Arguably the month's most-used release: on July 8, OpenAI shipped GPT-Live, a full-duplex — i.e. bidirectional — voice model that listens and speaks at the same time, ending the turn-based era of Advanced Voice Mode for the 150M+ people who use ChatGPT voice weekly. Instead of trading discrete messages, it processes input and output continuously, deciding many times a second whether to speak, keep listening, pause, interrupt, back-channel ("mhmm"), or call a tool — so you can talk over it and it adapts. The genuinely novel part isn't the duplex audio (Google's Gemini Live already does that) but the delegation architecture: GPT-Live handles the live conversation while handing hard questions to a background reasoning model — GPT-5.5 at launch, and whatever's newest thereafter — so its intelligence ceiling rises automatically. It ships as GPT-Live-1 (Go/Plus/Pro) and GPT-Live-1 mini (Free), with nine remastered voices and Instant/Medium/High reasoning levels. No public API yet (TechCrunch); developers building voice agents use GPT-Realtime-2.1 over WebRTC/WebSocket/SIP in the meantime.
• Anthropic — Fable 5 restored, Sonnet 5 launched. Claude Fable 5 returned to global availability on July 1 after a 19-day export-control pause. Sonnet 5 arrived at near-Opus-4.8 quality with introductory $2/$10 pricing through August 31 — though its new tokenizer can consume up to 35% more tokens, muddying the headline discount. Anthropic also published interpretability research on a "J-space" global workspace inside Claude.
• Google — Gemini Enterprise. At Cloud Next '26, Google reframed its pitch around fleets of governed agents, not a chatbot — its direct answer to ChatGPT Work. It also shipped Nano Banana 2 Lite and Gemini Omni Flash, and won the Siri deal (new Siri runs on Gemini) even as Apple sued OpenAI over trade secrets.
• SpaceXAI — Grok 4.5. xAI rebranded to SpaceXAI on July 6 and shipped Grok 4.5, a 1.5T MoE that scores 83.3% on Terminal-Bench 2.1 while using ~25% as many output tokens as Opus 4.8 on similar tasks.
• Meta — Muse Spark 1.1. Adds a 1M-token context window and computer-use across desktop, browser, and mobile, with parallel subagent delegation.
• Agentic work products. Anthropic's Claude Cowork (July 7) runs long email/calendar/file tasks that continue even when your device is offline; OpenAI's ChatGPT Work (July 9) fuses ChatGPT with Codex so non-technical users can build documents, sheets, and apps from approved data.

What actually changed this month

The through-line, echoed across the July recaps, is that the industry moved from "best model wins" to "best fit wins." Price, speed, license, hardware footprint, and censorship posture now matter as much as a top-line benchmark. Two open-weight labs — one Chinese, one American — put models at or adjacent to the frontier within a day of each other, and neither is trying to win purely on leaderboard score. Moonshot is selling ownable capability at premium prices; Thinking Machines is selling a customizable, auditable base you can align to your own values.

But the quieter lesson is about iron, not intelligence. Kimi K3's power is inseparable from its cost: a model you can legally download and self-host is still gated by a 650 GB-to-1.7 TB memory wall and Moonshot's own "64+ accelerator supernode" guidance. Inkling makes the same point more gently — its NVFP4 checkpoint halves the GPU count, and Inkling-Small hints that clever pre-training can shrink the hardware bill further. For most organizations the decision in H2 2026 isn't "open vs closed." It's "do we own the GPUs to make 'open' mean anything?" — and for the very largest open models, the honest answer is increasingly rent them, or route to the API.

If you're weighing self-hosting against managed inference for one of these models, that trade-off — throughput economics, warm-capacity cost, and when owning the weights actually pays off — is exactly what our model leaderboard and our connect tooling are built to help you reason about.

---

Sources

• Tom's Hardware — China's 2.8-trillion-parameter Kimi K3
• VentureBeat — Moonshot releases Kimi K3, largest open-source model ever
• Fortune — Markets experience new DeepSeek shock after Kimi K3 · Axios — China just erased America's AI lead
• Simon Willison — Kimi K3, and the pelican benchmark
• modemguides — Can You Run Kimi K3 Locally? The 2.8T Hardware Reality
• Thinking Machines Lab — Introducing Inkling · TechCrunch — Thinking Machines' first open model
• The Register — Former OpenAI CTO releases a frontier model that's actually open · Databricks — Inkling now on Databricks · VentureBeat — Inkling multimodal open source
• OpenAI — Introducing GPT-Live · TechCrunch — OpenAI's new voice models for live conversations
• ThursdAI — July 2026 AI releases · llm-stats — AI updates, July 2026

Benchmark figures for Kimi K3 are Moonshot-reported and not independently reproducible until the weights ship (promised by July 27, 2026); K3 memory and GPU footprints are community estimates scaled from the K2 family. Inkling figures are drawn from published weights and vendor documentation. As always, treat vendor-reported numbers as claims until the community reproduces them on the leaderboard.