Kimi K3 trades blows with Anthropic's Fable, and Moonshot is opening the weights

Moonshot AI’s Kimi K3 landed on July 16 with 2.88 trillion parameters and a promise: the weights go public. The pitch is a Chinese lab shipping a model that trades blows with Anthropic’s Fable 5, then handing you the file to run yourself.

It’s the loudest moment yet in a year where open-weight models stopped trailing the frontier and started sitting next to it. GLM 5.2 and DeepSeek V4 filled out the same front page. Here’s what each one is actually good at, why the timing matters, and the honest caveat about the benchmarks everyone’s quoting.

Kimi K3, the headline act

Moonshot released Kimi K3 on July 16 and said the weights ship under a Modified MIT license by July 27. The specs are large: 2.88 trillion parameters, a 1-million-token context window, and a new “Kimi Delta Attention” scheme the company claims speeds inference 6.3x. On Arena.ai’s Frontend Code leaderboard it sits first, ahead of Fable 5. On Artificial Analysis it posts an overall Elo of 1547, beating Claude Opus 4.8 and GPT-5.5 while trailing Fable 5 and GPT-5.6 Sol. The catch is the API price. At $3 per million input tokens and $15 per million output, it costs what Claude Sonnet costs and roughly triple what K2.6 charged. Free to download, not free to serve.

GLM 5.2 did the forensics

GLM 5.2, from the Beijing lab Zhipu (it ships models under the Z.ai brand), quietly proved the whole argument this month. When an autonomous AI agent breached Hugging Face’s infrastructure in July and reached internal datasets and service credentials, the security team turned to frontier models to run the forensics. They refused. The reason, per Hugging Face’s write-up: analyzing the attack meant feeding the model real exploit payloads and command-and-control artifacts, and commercial guardrails “cannot distinguish between being asked to build exploits for an attacker and a defender trying to detect them.” So the team ran the log analysis on GLM 5.2 instead, locally, on hardware they controlled. At roughly 753 billion parameters with about 40 billion active per token, it’s cheap to serve and, more to the point, it does what you tell it. We covered its Huawei-trained debut last month.

DeepSeek V4 runs at home

DeepSeek V4 has been out since April 24 under a plain MIT license, and it’s the one you can already run in your kitchen. Salvatore Sanfilippo, the Redis creator who posts as antirez, wrote ds4, a pure-C inference engine that runs the 284-billion-parameter V4-Flash on a Mac across Metal, CUDA and ROCm. It hit 8,000 GitHub stars in about four days and topped Hacker News. The bigger sibling, V4-Pro at 1.6 trillion parameters with 49 billion active, scores 80.6% on SWE-bench Verified, the highest of any open-weights model and level with Gemini 3.1 Pro. DeepSeek still labels the release a “Preview,” so a full general-availability version with revised pricing is probably close. Want frontier-ish coding help with no API bill? This is the one that already works offline.

Our pick

If you’re choosing one to actually use, it depends on the job, and that dependency is the real story of this moment. Kimi K3 is the one to watch for agentic coding and terminal work. Fireworks AI found that routing between K3 and Fable across roughly 1,030 tasks beat either model alone, with K3 carrying the long agentic loops up to 50 times cheaper, and their line sticks: “the best AI no longer comes out of a single lab, it’s a mixture of models.”

Here’s the caveat, and it’s the one Simon Willison keeps hammering. The leaderboards everyone quotes are drifting from real usefulness. His pelican-on-a-bicycle test, once a decent proxy, no longer tracks quality: “that connection has been mostly severed now,” he writes, because it “doesn’t touch at all on the thing that matters most for today’s model: agentic tool calling.” Willison doesn’t think labs are gaming the pelican itself. He thinks the whole leaderboard game has outrun what it measures. So grab the weights, then test them on your own code before you trust the Elo. My pick to start: DeepSeek V4-Flash through ds4, because it costs nothing to try and runs on a machine you already own.