Skip to content

Topic

Open-Source AI

The practice of publishing model weights, training code and datasets under terms that let others run, modify and redistribute them.

Current stories

buildOne report1 publisher

AWS open-sources a 2B model that picks from a fixed list of options, plus a policy sandbox

AWS open-sourced Strands Decider 2B, a 2B-parameter decision model it says makes a decision in a median of about 115 ms on an RTX 3090. Strands Box, the sandbox released with it, enforces agent rules in a kernel-level policy layer outside the model.

Publishers:infoq.com

Reality

Evidence40
Adoption
Not scored
Hype gap+10
Incentives
Not scored
Confidence50
buildOne report1 publisher

Ai2's Olmo-core 3 trains a 47B MoE at 2.7 times its old throughput under random routing

Ai2's Olmo-core 3 reached about 52,000 tokens per second per GPU on a 47B mixture-of-experts model, roughly 2.7 times the old stack, on eight B300 GPUs. The report labels every headline rate as measured under random routing, so a team with a trained router has to rerun the comparison.

Publishers:dev.to

Reality

Evidence55
Adoption
Not scored
Hype gap+10
Incentives
Not scored
Confidence58
productConfirmed4 publishers

Alibaba's $25bn a year of e-commerce cash backs Joe Tsai's open-source pitch to Europe

Alibaba chairman Joe Tsai pitched open-source AI to Europe as its only path to independence while Alibaba funds AI with about $25bn a year of e-commerce cash. Three years of doubled compute spending make Alibaba's AI push a strategy buyers can plan around, and Tsai's own warning about trusting foreign technology covers Alibaba too.

Perspective Coverage

4 publishers
Builder
Builder 30%
Operator
Operator 35%
Investor
Investor 35%

Reality

Evidence55
Adoption40
Hype gap+25
Incentives80
Confidence60
buildConfirmed3 publishers

DeepSeek packages its unaudited Harness agent runtime as a Windows and macOS desktop app

DeepSeek released its open-source Harness agent runtime as a Windows and macOS desktop app on September 30, so users no longer need a command line to run it. The agent edits local files and runs commands, so what each user exposes to it decides what it can touch and which provider sees it.

Perspective Coverage

3 publishers
Builder
Builder 43%
Operator
Operator 37%
Investor
Investor 20%

Reality

Evidence72
Adoption30
Hype gap+10
Incentives55
Confidence65
buildConfirmed2 publishers

Ai2's Olmo-core 3 gets 2.7x its old MoE throughput by keeping experts resident on GPUs

Ai2 released Olmo-core 3, an open mixture-of-experts training stack it has benchmarked at over one trillion total parameters. Its speedups are measured against Ai2's own earlier FSDP code, so teams on Megatron-Core must run their own comparison.

Perspective Coverage

3 publishers
Builder
Builder 55%
Operator
Operator 25%
Investor
Investor 20%

Reality

Evidence50
Adoption
Not scored
Hype gap+20
Incentives55
Confidence60
buildOne report1 publisher

Eclipse's Sovereign AI Foundation sets out to measure what AI provider lock-in costs

Eclipse Foundation launched the Sovereign AI Foundation with 17 mostly European members to help organizations manage dependence on a single AI provider. Its director lists where switching costs land and says members will now gather the evidence to size them.

Publishers:thenewstack.io

Reality

Evidence40
Adoption15
Hype gap+10
Incentives55
Confidence45
investOne report1 publisher

Bessent proposes a US-China AI incident channel in the week Trump rejected a 'globalist scheme'

Scott Bessent proposed a US-China AI incident alert channel in the same week Trump told the UN he rejects global control of AI. The channel asks nothing of labs before launch, while Xi's main ask, looser US chip export controls, is the item with revenue at stake.

Publishers:fortune.com

Reality

Evidence50
Adoption8
Hype gap+20
Incentives65
Confidence45
productOne report1 publisher

Meta says it built Muse from scratch while keeping OpenClaw's file names and personality lines

Meta's Muse, with an estimated 600,000 US daily users, reuses the core file names and personality lines of open-source agent OpenClaw. Meta says the code is its own, so the case for Muse depends on how it guards the logins its agent uses.

Publishers:theverge.com

Reality

Evidence45
Adoption40
Hype gap+35
Incentives70
Confidence45
investOne report1 publisher

Washington rejects the global AI rulebook OpenAI and Anthropic asked the Security Council to write

Dario Amodei and Sam Altman told the UN Security Council that AI could threaten humanity and asked governments to set safeguards. The White House science adviser in the room said no new global governance structure is needed.

Publishers:livemint.com

Reality

Evidence72
Adoption12
Hype gap+25
Incentives78
Confidence62
buildOne report1 publisher

Mila splits the open AI foundation layer into an interface standard and a reference install

Mozilla put up the first $5 million and Hypertec added $1 million for first-year Canadian deployments on its own hardware. The published contracts are the part that would let anyone swap the pieces.

Publishers:blog.mozilla.org

Reality

Evidence36
Adoption12
Hype gap+40
Incentives80
Confidence52
productOne report1 publisher

Hugging Face, a company Nvidia is buying for $12.93bn, asks to join Anthropic's third-party evaluator program

Anthropic and OpenAI have promised outside evaluators permanent, employee-level access to their systems. Neither has published the terms, and the first volunteer is mid-acquisition by a would-be anchor investor in Anthropic's IPO.

Publishers:thenextweb.com

Reality

Evidence56
Adoption18
Hype gap+38
Incentives82
Confidence58
buildOne report1 publisher

Xi puts China in charge of building a BRICS open-source AI community

Xi used the summit's closing session to make China the lead builder of an open-source AI ecosystem for BRICS members, and Beijing says it has already partnered with around 30 countries on a separate World AI Cooperation Organisation.

Reality

Evidence45
Adoption
Not scored
Hype gap+45
Incentives80
Confidence40