VC
Value Add VC
⚡HomePulse⚡Helpful Apps📝Blog
Illustration for: Moonshot's Kimi K3 Ships as Largest Open-Weight Model Ever
Value Add VC/Pulse/AI

Moonshot's Kimi K3 Ships as Largest Open-Weight Model Ever

Moonshot AI released Kimi K3's full open weights, a 2.8-trillion-parameter model described as the largest open-weight release in history, requiring roughly 1.4 terabytes of memory to run.

2.8 trillion
Parameters
1.05M tokens
Context window
~1.4 TB
Memory (MXFP4)
~5.6 TB
Memory (16-bit)
Modified MIT
License
TC
Trace Cohen
Early-stage VC & angel · Founder, New York Venture Partners
July 27, 2026
2 min read
ShareXLinkedInEmail

THE RUNDOWN

1

Kimi K3's full weights went live on Hugging Face at roughly 7:30pm ET on July 26, a day ahead of schedule, under a Modified MIT license with a 1,048,576-token context window

2

At 2.8 trillion parameters, the model needs about 1.4 terabytes of memory even compressed to MXFP4 four-bit precision, and 5.6 terabytes at full 16-bit precision

3

It's the third frontier-scale Chinese open-weight release in weeks, following Alibaba's Qwen3.8-Max preview and DeepSeek's V4 general availability

4

The release lands under an active, unproven White House accusation that Moonshot distilled Anthropic's Fable model, a claim researchers say lacks public evidence

TC

The VC Read · Trace's Take

Trace Cohen

A free 2.8-trillion-parameter model that needs 1.4 terabytes of memory isn't really 'open' for 99% of builders -- it just moves the gatekeeping from the model license to whoever can afford Together AI or Modal's hosting bill. The real story for founders isn't the download, it's who prices access to it first.

AI Landscape → AI Chip Wars →

Analysis

Moonshot AI released the full open weights of Kimi K3 to the public, making what Bloomberg describes as the largest open-weight AI model ever published freely downloadable, a day ahead of its original July 27 target. The 2.8-trillion-parameter model, licensed under a Modified MIT license, carries a 1,048,576-token context window and was published to Hugging Face at roughly 7:30 p.m. Eastern on July 26.

The download itself is a serious infrastructure commitment in its own right: even compressed into MXFP4 four-bit precision, the weights require roughly 1.4 terabytes of fast memory, ballooning to about 5.6 terabytes at full 16-bit precision -- before any context window is even loaded. Together AI and Modal both announced day-zero hosted access timed to the release, recognizing that most organizations lack the raw infrastructure to self-host a model this size even if the weights are free.

Kimi K3 lands in an increasingly crowded open-weight race that has scrambled Western AI labs' assumptions about where the frontier sits. Alibaba previewed its own 2.4-trillion-parameter Qwen3.8-Max just days earlier at the World AI Conference in Shanghai, and DeepSeek's V4 model reached general availability in July as well, meaning three separate Chinese labs have now shipped frontier-scale open or near-open models inside a matter of weeks -- a pace that has no real precedent among US labs, where OpenAI and Anthropic have kept their flagship models closed.

“Kimi K3 lands in an increasingly crowded open-weight race that has scrambled Western AI labs' assumptions about where the frontier sits.”

The release also lands squarely inside an active political fight: the White House has publicly accused Moonshot of covertly distilling Anthropic's Fable model to build Kimi K3, a claim AI researchers say lacks public evidence and that Moonshot disputes, while 25 US tech companies -- notably excluding OpenAI and Anthropic -- have lobbied Washington against new restrictions on open-weight models. Kimi K3 shipping as the largest open-weight release in history, under that cloud, makes the model as much a geopolitical Rorschach test as a technical achievement.

For builders, the practical calculus is now less about whether a capable open-weight frontier model exists and more about who can afford to run it: a 2.8-trillion-parameter model needing 1.4 terabytes of memory is out of reach for the vast majority of startups without hosted access from providers like Together AI or Modal, which effectively re-creates a gatekeeper layer even on a nominally "open" model.

What to watch: independent benchmark verification of Kimi K3's real-world performance against Claude Opus 5 and GPT-5.6 Sol, how quickly hosted inference pricing settles given the model's extreme memory requirements, and whether Washington's distillation accusation against Moonshot escalates into an actual policy action that would contradict the open-weight coalition's lobbying push.

ShareXLinkedInEmail

Analysis and editorial commentary by Value Add Pulse.

← Back to Pulse

THE WIRE in your inbox— Tech, startup & VC news with Trace's take. Free, no spam.

Read Next

AI· Jul 26, 2026

TechCrunch Makes Sense of the Chinese AI Panic

Illustration for: TechCrunch Makes Sense of the Chinese AI Panic
AI

TechCrunch Makes Sense of the Chinese AI Panic

A TechCrunch analysis argues the panic following Moonshot's Kimi K3 release reveals more about American AI-competitiveness anxiety than about the model itself.

AI· Jul 27, 2026

Jensen Huang: Chips Must Grow 10x for AI Agents

Illustration for: Jensen Huang: Chips Must Grow 10x for AI Agents
AI

Jensen Huang: Chips Must Grow 10x for AI Agents

Nvidia CEO Jensen Huang told Bloomberg the semiconductor industry must grow roughly tenfold over the next decade to serve computing dominated by autonomous AI agents.

AI· Jul 26, 2026

$7 Trillion AI Buildout Threatens Cheap Power

Illustration for: $7 Trillion AI Buildout Threatens Cheap Power
AI$7 Trillion Buildout Risk

$7 Trillion AI Buildout Threatens Cheap Power

A Fortune analysis warns that a $7 trillion AI data center buildout with no guaranteed matching demand threatens to reverse years of falling US electricity costs.

@Trace_Cohen·t@nyvp.com