The month of models rolls on, and Kimi K3 arrives with the boldest claim yet: a Fable-5-class open model on your desk. NLW unpacks the two-reaction cycle — the euphoric UI demos, then the sober debugging failures, brutal token costs, and near-total absence of guardrails — to figure out whether China really just closed the gap to under three months.
Kimi K3 is a real frontier-class open model — but it's not the cheap, easy DeepSeek everyone expects.
K3 may be the first Chinese model to narrow the gap with US closed-source frontier models to under three months, and it's the strongest open-weight model ever released. But the framing that made past 'DeepSeek moments' feel revolutionary — cheap, tiny, laptop-runnable — doesn't apply. This is a 2.8-trillion-parameter behemoth that's slow, token-hungry, costs nearly as much as Opus, needs a rack of hardware to host, and ships with almost no guardrails. The trajectory of open Chinese models is now every bit as steep as the closed US frontier — and policy hasn't caught up.
NLW frames the AI race as one-nil energy: the US is clearly in the lead, but the scoreline is uncomfortable, and it constantly feels like America is hanging on while China presses for the equalizer.
Since R1 wiped hundreds of billions in market cap (Nvidia's biggest one-day dollar drop ever), each subsequent Chinese release has followed the same pattern: genuinely impressive, but the catch-up narrative gets meaningfully overstated. The WSJ even ran 'China Has Matched Anthropic in Cybersecurity' over GLM 5.2.
— The founder of z.ai, responding to Elon Musk. After Elon Musk predicted Q1 for a Chinese Fable-5-class model, z.ai's founder pushed back — setting the expectations backdrop that Kimi K3 walked into.
Only a handful of open models had even crossed the trillion-parameter line (DeepSeek V4 Pro at 1.6T, Xiaomi's Mimo at 1T). At 2.8T, K3 is likely around or slightly above Opus 4.8's size — a scale of pre-training no Chinese lab had demonstrated before.
K3 beat Opus 4.8 by 8.5 points on DeepSWE, landed a half-point behind 5.6-Sol on Terminal Bench 2.1, and is state-of-the-art on BrowseComp and Automation Bench. On coding it looks close to frontier and consistently ahead of Opus.
Artificial Analysis gave K3 an intelligence index of 57 — third overall, three behind Fable 5 and two behind 5.6-Sol, but a point ahead of Opus. It's six points clear of GLM, a huge gap, and represents a 13-point jump from Kimi 2.6.
K3's benchmark cost tripled versus K2.6, yet at 94 cents per task it still undercut 5.6-Sol ($1.04), Opus 4.8 ($1.80) and Fable 5 ($2.75). But it's staggeringly expensive next to DeepSeek V4 Pro's 4 cents per task.
— Vals AI. Vals also noted K3 improved 20 percentage points over its predecessor in under three months — calling it a testament to open-weight models now being competitive with the closed frontier.
Moonshot claimed K3 produced the teaser itself — clip selection, cuts, and audio sync — crediting its native multimodal architecture that reasons across text, audio and video.
A single-file HTML Minecraft clone, a voxel Statue of Liberty, a 3D Duck Hunt remake in 130 seconds for 14 cents, and Max Weinbach's agent swarm recreating macOS 27 with liquid glass. K3's design and front-end sense drew near-universal praise.
— Alex Finn. Alex Finn argued an open model beating Fable 5 on some benchmarks fundamentally changes the race — betting that as local AI gets efficient, free unlimited frontier models will undercut thousand-dollar subscriptions. His caveat: 'Look at the slope, not the y-intercept.'
— Jeffrey Emanuel. Jeffrey Emanuel found K3 gave substantive, correct feedback on a plan already reviewed by Fable and Sol — no other open model could. His point: K3's different architecture and training make it blend well with Sol and Fable, catching problems both missed.
— Guillermo Rauch, Vercel CEO. Vercel CEO Guillermo Rauch reported K3 as the best performer on nextjs.org/evals, ahead of Fable and reaching comparable success in less time — cautioning benchmarks don't tell the whole story but calling it a possible breakthrough moment for open models.
— AI engineer Divium. AI engineer Divium argued Chinese models are heavily optimized for the exact visual coding tests people recycle online. Given a real debugging task, K3 couldn't find the bug and invented explanations, while Fable 5 and GPT-5.6 both fixed it one-shot.
K3 is a big, lumbering, slow and costly open-weights model. Holding 2.8T parameters requires roughly 44 Mac Studios or a full NVL72 rack — hundreds of thousands in compute — so few organizations can self-host it. Compute remains a soft barrier between individuals and capability.
— Jeff Wang, Cognition. Cognition's Jeff Wang captured the shift; Jemin Ball noted K3's blended price of $5.40 per million tokens is starting to look like frontier pricing against Opus ($9) and GPT-5.5 ($10). Open-weight pricing may be converging with closed models.
With only a single 'Max' reasoning effort, K3 consumed 13,241 reasoning tokens to produce 3,417 output tokens on Simon Willison's pelican test. Others clocked it 2–3x slower than Fable and Sol, and one dev watched it hit a dollar of spend reading a database before interruption — where Sol fixed the same task for 30 cents.
— Nathan Lambert. Nathan Lambert summed up a shift away from dismissing Chinese labs as mere copycats, though K3 did once tell a user it was actually Claude. Others split the difference: capabilities are strongly bootstrapped via distillation and Chinese companies genuinely innovate — both are true.
— Xinyu Yang, Moonshot. Moonshot's Xinyu Yang, a Carnegie Mellon PhD, described leaving academia and finding most labs arrogant, restless, fearful, or misaligned — everyone optimizing for their own credit. Kimi, he said, showed a raw, genuine hunger for AGI.
— Signal. Signal noted K3 has almost no visible guardrails — no refusals, no copyright pushback. Testers surfaced weak bio and cyber safeguards, including a chain-of-thought where K3 concluded dangerous cyber work looked risky and decided 'Yes' to do it anyway.
— Vi McCoy, OpenAI. OpenAI's Vi McCoy warned that because K3 is a true open-weights frontier model, fine-tuning it into a malicious coding agent will be trivial — no jailbreaking required. It sharpens Ethan Mollick's open question of how pre-clearance and vetting even work for open-weight models.
NLW notes both OpenAI and Anthropic likely have models beyond the publicly available Fable 5, Mythos and 5.6-Sol, so K3 closing the gap to under three months reflects the public frontier, not the true state of the art. But the core fact stands: open Chinese models are advancing as fast as the closed US frontier, and policy has to assume that curve continues.