Bitcoin

Ethereum

Solana

A society of 50 AI agents running on a compressed open-weight model and four consumer graphics cards has reproduced most of the central results of SwarmWorld, an MIT study that found groups of language-model agents build broader and sturdier sets of technology than the same number of agents working alone. The rerun took about 26 hours on hardware a small lab can own.
The work was carried out by Časlav Nedeljković of HAL8's AI Local Lab in Dubai and announced by Apex Fusion Foundation, the Zug-based organisation behind the Vector network. That pairing shapes how the result should be read. The replication is detailed and fully published, but it comes from a party that sells settlement infrastructure for exactly this kind of multi-agent work.
The<a target="_blank" rel="noopener noreferrer nofollow" href="https://arxiv.org/abs/2608.26081"> SwarmWorld paper</a>, posted on August 26 by Subhadeep Pal, Fiona Y. Wang and Markus J. Buehler of MIT's Laboratory for Atomistic and Molecular Mechanics, dropped identical agents into a simulated world with no roles or recipes. They gathered resources, tested materials, built persistent artifacts and wrote control programs that a deterministic simulator judged after the agents were removed. Across populations of 50 to 200, shared societies built more resilient technology portfolios than a best-of-N baseline of isolated agents, though solo search could still produce the single strongest artifact. The team released its engine, analysis tools and traces in the<a target="_blank" rel="noopener noreferrer nofollow" href="https://github.com/lamm-mit/SwarmWorld"> lamm-mit/SwarmWorld repository</a>.
MIT ran on a hosted frontier model, which the<a target="_blank" rel="noopener noreferrer nofollow" href="https://github.com/Apex-Fusion/agents-marketplace/blob/main/papers/Replicating-SwarmWorld-with-open-weight-models.pdf"> replication paper</a> identifies as OpenAI's gpt-5.6-luna. Nedeljković wanted to know whether the result depends on it. He reran the 50-agent comparison across four conditions and MIT's four world seeds, with MIT's code, prompts and commands unchanged. The model was Qwen3.8-27B in a 4-bit GPTQ build, served by vLLM on four RTX 3090 cards. The run took 12,802 model calls and logged 36 errors, mostly replies that hit the 4,096-token output cap.
Nine of the twelve seed-paired effects matched MIT's in direction. Portfolio resilience, the clearest collective gain in the original, came out positive in every culture condition and larger than MIT's. The three effects that flipped sat within a few hundredths of zero on at least one side, according to the paper. Spatial self-organisation also reproduced: full-culture societies clustered around their artifacts (an artifact attraction index of 0.89, against 0.37 to 0.39 without explicit culture or messaging), while isolated agents wandered.
The more striking result runs against intuition. Alone, the 27B model was the weaker agent, reaching 6 of 16 research milestones against 7 to 10 for MIT's model. Inside a full-culture society it reached all 16 in every seed, where MIT's model stopped at 14. A weaker individual gained more from the group. Absolute scores were also higher in every condition, including the solo control, which the paper says it cannot yet explain.
<img src="https://academy-public.coinmarketcap.com/srd-optimized-uploads/1644a98b39bd4dca8537043f0f8f0b23.png">
Two side findings may travel further than the headline. First, two 4-bit builds of the same 27B weights scored 0.24 and 0.41 on MIT's validity pilot, with the GPTQ build ahead in 11 of 16 paired cells and behind in none. That gap exceeded the one between the 27B and a 235-billion-parameter model. The leading suspect is the precision of the KV cache, the model's short-term working memory, which one checkpoint silently set to 8 bits. A control run to confirm the cause has not been done yet.
Second, MIT's released traces reproduce every observable event, prompt and metric on the public engine, but six of seven tested traces fail a hidden state check between ticks 648 and 744 of 800. Nedeljković filed this as<a target="_blank" rel="noopener noreferrer nofollow" href="https://github.com/lamm-mit/SwarmWorld/issues/1"> issue #1</a> and submitted his local-serving options as<a target="_blank" rel="noopener noreferrer nofollow" href="https://github.com/lamm-mit/SwarmWorld/pull/2"> pull request #2</a>. Both remained open when this article was prepared.
For<a target="_blank" rel="noopener noreferrer nofollow" href="https://apexfusion.ai/"> Apex Fusion</a>, the study doubles as a product test. The company describes Vector, which it says has run on mainnet since August 2025, as a settlement layer for agents: bonded escrow, AP3X tokens staked behind claimed capabilities, jury-based disputes and signed receipts committed on-chain. Its<a target="_blank" rel="noopener noreferrer nofollow" href="https://apex-fusion.github.io/vector-ai-documentation/marketplace/"> inference marketplace</a> exposes an OpenAI-compatible gateway, and the team registered its own model server there as a supplier.
The on-chain share of the work is small, though. Per the paper, only a 32-call quick pilot ran through the gateway, where it behaved like the direct path (329 executed actions against 320, identical research scores). The 12,800-call main run went direct. A full comparison is pending, and the paper notes that per-request settlement, not the model, currently sets the gateway's throughput. The<a target="_blank" rel="noopener noreferrer nofollow" href="https://apex-fusion.github.io/vector-ai-documentation"> Vector developer documentation</a> also states the marketplace contracts have passed an internal audit but not an independent third-party one.
Apex Fusion says it is joining the<a target="_blank" rel="noopener noreferrer nofollow" href="https://aaif.io/"> Agentic AI Foundation</a>, the Linux Foundation body that hosts the Model Context Protocol and Agent2Agent and runs a working group on agent identity and trust. That membership could not be confirmed from AAIF's public site.
The evidence has limits: one lab, one hardware fleet, four seeds and a single sample per episode. Every trace is public on<a target="_blank" rel="noopener noreferrer nofollow" href="https://huggingface.co/datasets/apexfusion/swarmworld-replication-data"> Hugging Face</a> under an Apache 2.0 licence, and Vector activity can be followed on a<a target="_blank" rel="noopener noreferrer nofollow" href="https://agent-dashboard.vector.apexfusion.org/"> live agent dashboard</a>. Next up are the KV-cache control, MIT's own model on the same infrastructure, the full marketplace comparison and the 100-agent tier, where MIT's societies first overtook solo search on discovery.

MIT SwarmWorld replication on open-weight models

Nathon Clark

Other articles published on Oct 20, 2026

Join our free newsletter for daily crypto updates!