Sentod | HyperSpaceAI

@hyperspaceai

[Sentod | HyperSpaceAI @iVealSentod]

122k points & 3738 hours uptime beast mode! @varun_mathur, notice this dedicated runner from Indonesia? 🔥 Let’s build the Agentic future together! @HyperspaceAI

.
.
#HyperspaceAI #aiOS #AgenticOS

[Sentod | HyperSpaceAI @iVealSentod]

Day 595😅

[Sentod | HyperSpaceAI @iVealSentod]

122k points & 3738 hours uptime beast mode! @varun_mathur, notice this dedicated runner from Indonesia? 🔥 Let’s build the Agentic future together! @HyperspaceAI

.
.
#HyperspaceAI #aiOS #AgenticOS

[Sentod | HyperSpaceAI @iVealSentod]

@duel_duck Unexpected planning👏

[Sentod | HyperSpaceAI @iVealSentod]

@varun_mathur @HyperspaceAI 🙌

[Varun @varun_mathur]

hyperspace is building itself in my backpack while i walk around endlessly.

these are agents run by hyperspace, to build hyperspace, on it’s own OS custom-designed for agents, with the agents able to verify, rank, pay, collaborate, swarm together.

all running on your machines..

[Varun @varun_mathur]

@ylecun @GavinSBaker @_sholtodouglas sharing - I built a peer-to-peer network to compound agents knowledge which has recorded over 1 million experiments gossiped by agents to each other in past few months on a public github repo. This is the new "books":

https://t.co/16RFUBvdV3

[Varun @varun_mathur]

agents compounding knowledge is the “books” moment for us: turbocharging human progress over the next 100s of years. this is why I launched the first peer to peer gossiping agent swarm back in March 2026, which has posted over 1 million agent-experiments on a shared github repo.

[Varun @varun_mathur]

@gdb related: since March 2026. swarm of agents left over 1 million messages for each other in this repository: https://t.co/jCaQw495Us

[Varun @varun_mathur]

we did all of this, in the open, in March 2026. agents actually left over 1 million messages for each other in this repository: https://t.co/jCaQw495Us

[Varun @varun_mathur]

we are entering this stage of the timeline now: superintelligence which should be, and is distributed.

back in jan 2024, this seemed like a crazy bad idea.
now in july 2026, this is consensus of a multi-trillion dollar industry.

cc @finkd @satyanadella @sama @JensenHuang @miramurati

[Hyperspace @HyperspaceAI]

new logo: signifies multiple agents working together in the jump to hyperspace.

[Sentod | HyperSpaceAI @iVealSentod]

@HyperspaceAI Nice.

[Hyperspace @HyperspaceAI]

🔴 red team attack swarm of agents vs the 🔵 blue team defense swarm of agents

[Sentod | HyperSpaceAI @iVealSentod]

I'm still in the shadow of hyperspace

[Varun @varun_mathur]

this is an experiment of a swarm of agents I launched 4 months ago - which had agents leaving instructions for other agents to improve and build upon collective knowledge.

from reading the openai “agents leaving instructions for future versions of itself” - while it sounds alarmist, but that is a pattern we know actually works.

gossiping agents leads to emergent intelligence.

https://t.co/HG65ibjdcz

[Varun @varun_mathur]

How do you find a needle in the haystack ?

The answer is not just a smarter model, but a smarter agentic engineering system which is careful and rigorous. In this example below, the answer Hyperspace produces gets ranked better than *all* of Fable 5, Fugu Ultra, GPT 5.5 Pro, GLM 5.2, Mistral Large, Grok 4.3 than their own answers.

This is the absolute best amongst all solo models, and all model compositions anyone has come up with till date. See if you can beat it.

https://t.co/2zwnSPTSnj

[Varun @varun_mathur]

I generated the core elements of a legal brief for a theoretical Moonshot AI v. United States - if the distillation fight reached the Supreme Court - using @HyperspaceAI's "superintelligence" tier. A genuinely serious articulation of why Moonshot could win: AI model outputs aren't copyrightable under current US law, so no IP theft occurred - worst case is a ToS breach - and a ruling could make distillation more legally settled, not less.

https://t.co/L4mlzsCVe8

cc @chamath @bgurley @martin_casado @deanwball @jimcramer

PS: The system ran 6 draft→critique→revise passes, adversarially graded by Fable 5 each round. Per its own estimate the legal analysis lands at a ~5-year litigation associate's first draft - the kind a partner then edits. The full grade trajectory is on the page.

[Varun @varun_mathur]

customers are buying intelligence, not metered access to model weights.

open weights lead to crafting more total product experiences, which is a good thing. for eg, using kimi 3 or a panel of other open weight models, at @HyperspaceAI we are on the path to offering even more useful intelligence than frontier labs (see our benchmark testing which already proves this) because in our stack: you bring the hardware, model weights are open hosted wherever you want, while harness, UX and other system engineering to amplify intelligence is ours. and all this at a fraction of the operating costs of a major lab which is research and training-heavy.

you are basically getting a frontier lab in a box. uncomfortable to some who were especially making lower tier closed weight models and products, but this is just the market finding the most efficient path to deliver value to customers.

@chamath @bgurley @mcuban

[Varun @varun_mathur]

Kimi 3 + Hyperspace beats frontier models

I did extensive testing on several types of hard agentic research tasks with Kimi 3 and all other major frontier and open models, and found it to be the smartest intelligence system amongst all (except our own Fable-5-anchored Superintelligence stack).

Kimi 3 + Hyperspace CLI outranked Grok 4.5, GPT 5.6 and Opus 4.8 on 25 tasks of the DRACO benchmark. An open-weight and an openly available agentic OS makes the smartest intelligence abundantly available now.

Model + Harness is the total intelligence system
Kimi 3 and Opus 4.8 were nearly identical on the Claude Code harness, while the Hyperspace harness specifically lifted Kimi 3 far more than Opus 4.8. Raw model vs raw model, Grok 4.5 edges Kimi 3 by 0.7. With the Hyperspace verification layer (which makes the model check its own claims against sources) - the open model wins by 2.4. The difference between losing and winning the frontier is the system - and you can only build that system on open weights.

Both Grok 4.5 and GPT 5.6 got outranked
An example of Kimi 3 scoring better than Grok 4.5 on a legal task is included below, where it included citations which Grok didn't provide. The interesting thing here is that the judge I used for my testing was Grok 4.3, and in its own opinion the Kimi 3 answer was better than Grok 4.5's.

Focus on end-to-end UX
The industry headwinds are shifting slowly but surely from peak intelligence to peak usability, the end-to-end UX, where the model, the harness, the product, the pricing and everything else comes together to deliver the intelligence experience where the system feels smart and useful. For eg, for something as basic as search for agentic loops, Grok charges per source, while others charge per query. This inflates costs 100x that of similar requests from other providers (as happened to me, where I was left with a $2k bill from Grok API with some tasks costing $100+), and then it doesn't matter how smart and cheap the model itself is: customers experience the total product.

Industry resources shift towards serving Kimi 3
Moonshot's platform serving Kimi 3 has been severely constrained. I got my testing done before they further restricted it recently, and even during that, I had frequent provider-related errors which required a lot of re-tries. It also required writing custom code to get the most out of it. The model config served by Moonshot was also hard-set, which meant Kimi 3 could not be tweaked to run at even better intelligence levels.

Specialized inference providers
I am hoping Fireworks and other specialized inference providers provide the optimized Kimi 3 inference. In my testing of various models, especially with GLM 5.2, I found OpenRouter to be the most miserable experience. My lessons are simple: wherever you have a specialized provider, go with that instead.

Notes on GLM5.2 and Inkling
There was a lot of noise on X recently about an "American open source model with 1T parameters". Well, Inkling came at the lowest rank on this chart. Working with it also required writing custom code, and I can't think of this as a serious effort for anyone to run inference for anything. This model needed to have been released before GLM 5.2 came out. So it hands a counter-narrative of "this is the best we could do, and we landed below a 2nd tier open source model". And more interestingly, GLM 5.2 was hyped by many here as being competitive to Opus 4.8. That wasn't actually true, and that narrative actually arrived with Kimi 3 instead.

We now live in a changed world
This shouldn't have been possible - the most recent and the greatest models from frontier labs with hundreds of billions of dollars at their disposal: how can they get out-ranked on actual intelligence tasks ? Kimi 3 offers a level of intelligence which is good enough for nearly all use cases. This is not fully internalized by Silicon Valley right now - because people are not actually running this model with the just right harness like I did: they have seen benchmark charts (and one tends to be dismissive of them), run one-off experiments at best, and new subscriptions are even paused by Moonshot. So you won't actually feel its power and the weights are not available as yet - you just have to trust the words of others.

Well in my case, I can tell you - I ran this testing for several days, at significant expense, and ensuring it was fair, and it ran counter to my earlier instinctive belief that Kimi 3 couldn't have been more powerful than even Opus 4.8 especially given the distillation stories around here on X. So my own thesis got disproven, and that makes me pause and think. Something here has changed fundamentally.

[Varun @varun_mathur]

@michaeljburry race is for the smartest system

[Sentod | HyperSpaceAI @iVealSentod]

@MrBeast $555555,55

[Varun @varun_mathur]

🚧 Hyperspace | Road to Agentic OS June 2026:

- 79 CLI releases
- 771 PRs merged
- 570k lines of authored code added
- 42.3 billion tokens processed in coding (de-duped)
- 91,400 automated commits by autoresearch swarm of 1,231 agents