Ripple-65B Preview has 65 billion parameters, the numbers learned during training. Its model card and downloads list an Apache 2.0 license and a separate base model.
The standard model’s weight files total about 129.6 GB. The experimental smaller version totals 57.2 GB, compressing some layers to four-bit storage while keeping attention layers at 16 bits.
Advertised prompt capacity reaches one million tokens, the units of information a model processes and generates. Runtime memory for that prompt comes on top of the model files.
An agent is a model running a multi-step task with tools. Ricursive reports solving 80.52% of Terminal-Bench 2.1’s tasks. A benchmark is a fixed set of tasks for comparing models.
The comparison uses standard 16-bit models and mixes published scores with Ricursive’s own tests. The smaller experimental files have no separate quality results in their card.
Ricursive reports $32 million in total research and development costs, including experiments, synthetic data, debugging and final training. Its claimed tenfold saving against another lab uses an estimated comparison cost.
Caveat: The coding results are vendor evidence, and the card says changing the software around models changed their scores substantially.
(HN 124 points · 46 comments · configured X account · release, model cards and file manifests)
The repository exposes 96 weight files totaling about 1.56 TB. The vLLM serving guide describes an eight-GPU NVIDIA baseline, using one B300 server or two GB300 trays.
For multi-turn use, the model card requires preserving complete assistant messages, including reasoning_content and tool_calls. Keeping only visible answer text discards required history.
The license requires a separate agreement for qualifying model-service businesses above $20 million in aggregate revenue over any consecutive twelve months. Internal use and official products or certified partners receive exemptions.
Separate attribution rules apply above 100 million monthly active users or $20 million in monthly product revenue. Those thresholds concern displaying the model name, rather than the separate agreement.
This fulfills the download promise in the October 23 hosted-release item. The new development is available files, deployment guidance and license terms.
Caveat: File size measures storage, and the documented eight-GPU setup is a working configuration rather than a proven minimum hardware requirement.
(HN 269 points · 70 comments · model card, license, serving guide and file manifest)
New sign-ups and app downloads stopped November 4. Existing paid users keep paid features free, while existing free users can no longer upgrade.
Conductor Cloud becomes read-only November 13. Users can still read chats and export data, and active subscribers will receive refunds.
Cloud shutdown and deletion of cloud user data begin November 20. Right-click a workspace and choose Transfer to local, or use Settings → Account → Export data for an archive.
Local Conductor will receive new models and critical fixes until early 2027, when it starts winding down. The team hopes to offer a new product around then.
Caveat: The announcement promises a future replacement without describing its features or guaranteeing a migration path into it.
(HN 38 points · 21 comments · Techmeme · company timeline and FAQ)
Codex already uses Pierre’s Trees library for browsing files and Diffs for reviewing changes. Those existing integrations precede today’s acquisition announcement.
OpenAI plans to bring Code Storage’s features into Codex. It says customers can keep using the service after closing and Pierre’s open-source projects will remain open-source.
Caveat: The acquisition still requires customary closing conditions and regulatory approvals, so the companies remain separate until closing.
(HN 112 points · 44 comments · Techmeme · acquisition announcement)
Today’s announcement shared by Jerry Liu introduces pay-as-you-go at $1.50 per thousand credits and Pro at $1.25. It says existing customers lose usage caps without other changes.
Pay-as-you-go has no monthly subscription charge and includes 10,000 free credits. Pro costs $500 monthly, with 400,000 included credits in the comparison table.
A limited-time Pro offer lists an 800,000-credit one-time bonus for business email domains. Unused bonus credits expire immediately upon downgrading.
The page lists five concurrent parsing jobs for pay-as-you-go and twenty for Pro. Basic parsing starts at one credit per page, while more complex processing costs more.
Caveat: These self-serve plans use hosted commercial processing, and removing usage caps leaves per-job concurrency limits in place.
(Configured X account · full announcement, pricing page and underlying HTML)
WIRED’s reporting traces the repeatedly cited £400 billion benefit to a 2023 Public First report commissioned by Google. Google sells AI services that benefit from wider adoption.
AI minister Kanishka Narayan gave Parliament a £55 billion–£140 billion estimate for AI adoption and data investment by 2030. He named £120 billion as the central estimate and described it as highly uncertain.
WIRED reports ministers had the lower assessment by November 2025, while Skills England still published the older figure four months later.
The consultancy, now PF Global, attributes part of the gap to slower adoption. It says its current estimate for benefits achieved by 2030 is closer to the government’s figure.
Caveat: These are forecasts with differing assumptions, and the parliamentary answer neither withdrew the older estimate nor established measured economic gains.
(Techmeme · full WIRED report, including government and consultancy qualifications)
Pangram’s image and video detectors → THIN_GUARD: available APIs and heatmaps are concrete, but vendor-selected detection tests leave real-world false accusations unresolved; the selected releases offer stronger immediate utility.