Dark theme
WELCOME TO QQUANTUM

Let Qubi show you around.

Choose how you’d like to explore.

Don’t agreeGo to yourOS instead ↗

Both choices allow Google Analytics, so we can see which pages help.

Sound is your choice. Essential cookies keep the site working. Continuing either way also allows Google Analytics (visits and pages viewed).

Read our Terms, Privacy and Cookies.

Change your choices anytime from the footer.

Getting Qubi ready.
Engineering notes 22 November 2026 By QQuantum.AI 7 minute read
Open Weights Agentic Coding On Premise Ai Ai Evals Model Routing

Ripple-65B Preview: an Open Coding Model You Can Self-Host

Ricursive Intelligence says its Apache 2.0, 65B dense Ripple-65B Preview scores 80.52 on Terminal-Bench 2.1 and fits on one DGX Spark. The scores are self-reported. What it would take for a European team to trust it on-premise.

AI-drafted, machine-checked against our published editorial rules, and reviewed on publication. Corrections: [email protected] How we use AI

Qubi, a blue glass orb with two orbiting electrons, guides a floating, glowing crate through a lit doorway into a dark Barcelona server room, where a pink client orb with round glasses waits beside a locked rack.

Article

Ricursive Intelligence has released Ripple-65B Preview, a 65-billion-parameter dense model under Apache 2.0 that it says scores 80.52 on Terminal-Bench 2.1 and fits on a single NVIDIA DGX Spark. If the claims hold, a capable coding agent can run inside a European company's own network. Every score is self-reported, so the right response is a structured internal test, not a migration.

The release is described on Ricursive's announcement page and in the model card on Hugging Face, both dated 6 November 2026 in their page metadata; the announcement gives no exact release date in its text. We researched this piece on the same day and found no independent coverage or reproduction of the results. Everything in the next two sections is Ricursive's own account.

What Ricursive released

Ripple-65B Preview is, in the words of the model card, "a 65B parameter dense Transformer model". The weights are published under the Apache 2.0 licence, and the card states "support for up to 1M context length".

The announcement explains that the team "chose to train a modernized dense Transformer model rather than use Mixture-of-Experts (MoE)". A Mixture-of-Experts model routes each token through a subset of its weights; a dense model uses all of them every time. Ricursive describes the shape as a "narrow d_model (5120) and 80 total layers", so the model is deep rather than wide.

Two background details matter for provenance. Ricursive says the model is "a general-purpose variant of the proprietary model that powers our AI chip design platform", and that it "opted not to buy any posttraining data". It also puts the cost of developing and training the model at $32M, a figure that is the company's own and has not been audited.

The model card is candid about what "Preview" means: "Our Preview release has gone through relatively light posttraining, with a focus on agentic coding." It also describes the pretraining data as a "predominantly-English and code corpus".

Qubi, a blue glass orb with two orbiting electrons, hovers beside a single compact workstation on a desk in a dark Barcelona office, cyan light from a terminal window falling across a stack of server drives.

The headline number, and what it is compared against

The headline claim on the announcement page is a "U.S. open-weight" best of 80.52 on Terminal-Bench 2.1, a benchmark of agentic tasks carried out in a command-line environment. On Ricursive's own comparison chart, Reflection Beam 501B scores 80.1 and Meta's Muse Spark Open 816B scores 76.4. We have not checked those comparison figures against the other labs' own sources.

Three qualifications sit next to that number.

First, "U.S. open-weight" is a deliberately narrow category. It leaves out the Chinese open-weight families, such as Qwen, DeepSeek, GLM and Kimi, which a European buyer can also download and self-host. The claim is not that Ripple leads all open-weight models, and it should not be repeated as if it were.

Second, the lead over the next model on the chart is under half a point. On an agentic benchmark, where results vary from run to run, that is a tie for practical purposes.

Third, the model card says itself that "Replicating benchmark scores has been challenging, particularly for modern agentic evals", and notes that some values in its table come from in-house evaluations rather than official model cards. The comparisons are therefore not like-for-like. That candour is welcome, and it is also the strongest reason to run your own tests.

Why a dense 65B matters for on-premise teams

This is our analysis, not Ricursive's. The interesting part of the release is not the benchmark rank but the shape of the model.

Large open-weight coding models have tended to be big Mixture-of-Experts systems. They can be efficient per token, but serving them well means holding every expert in memory, and often splitting the model across several GPUs with the routing and parallelism work that implies. A dense 65B model is a much simpler serving problem: one set of weights, predictable memory use, and well-understood inference tooling. Ricursive says the model "can fit on consumer hardware such as the NVIDIA DGX Spark", and reports throughput of roughly 250 tokens per second on two NVIDIA B200 GPUs before quantisation. It published no throughput figure for the DGX Spark itself, so "fits" should be read as "loads", not "runs fast enough for a team". Our GPU sizing guide covers how to turn that distinction into a hardware decision.

For an enterprise team in the EU, that changes the conversation. Coding agents read source code, configuration and sometimes secrets. Many organisations have so far had two options: send that material to a frontier API, or accept a much weaker local model. A credible model of this size on a single workstation-class machine opens a third option, an on-premise or even air-gapped coding agent with the code never leaving the building. That speaks directly to data residency requirements.

The licence helps. Apache 2.0 is a standard permissive licence that legal teams already know, without the bespoke acceptable-use terms that some open-weight releases carry. That shortens procurement, though it does not remove the need for a model supply-chain review.

On cost, the trade-off is qualitative until the model has been measured. Self-hosting turns cost per task from a metered token charge into fixed hardware plus the engineering time to operate it. That suits steady, high-volume agent work on sensitive repositories. It suits occasional use badly, because the machine sits idle and the frontier API would have been cheaper and better. Our page on open weights versus proprietary models sets out that decision in more detail.

What to check before trusting it

A self-reported preview from a lab that is new to most buyers is a candidate, not a dependency. Before any production use we would check four things.

Your own repositories. Build a small evaluation set from real, closed tickets: the task description as input, the merged change and its tests as the reference. Run the agent in a sandbox, let it edit and run the tests, and score whether the tests pass and whether a reviewer would accept the diff. Our AI evals guide describes how to build that kind of test rig. This is the only number that predicts behaviour on your code.

Spanish and Catalan. The corpus is predominantly English and code. Code itself is largely language-neutral, but tickets, comments, commit messages and documentation in Barcelona teams often are not. Include tasks written in Spanish and Catalan and check that the agent reads them correctly and replies in the expected language.

Behaviour beyond the benchmark. Light post-training means refusal behaviour, guardrails and tool-call reliability are not documented. Test what the agent does when a command fails, when it is asked to delete files, and when a prompt injection sits inside a README. Wrap it in the same permission limits and review gates as any other agent.

Provenance. Record where the weights came from, verify checksums, and note that the model descends from a proprietary chip-design model whose training data is not described in detail. Keep the weights pinned to a reviewed version rather than pulling updates automatically.

Qubi and a green team-member orb hover over a long workbench of small sandboxed terminals, Qubi watching one screen glow cyan while a warm desk lamp lights a tray of test cards being sorted into two piles.

Where it fits in a routed stack

Even if Ripple-65B Preview passes those checks, we would not make it the only model. A more robust design is hybrid model routing: a local agentic coding tier handles work on sensitive repositories and routine changes, and harder or less sensitive tasks are routed to a frontier model through an API, with the routing rules written down and auditable. If the local model degrades or a better open model appears next quarter, it is swapped behind the router without changing the tools developers use.

A worked example, clearly hypothetical: an internal platform team gives the local tier dependency upgrades and test fixes on its private monorepo. The agent opens a branch, makes the change, runs the test suite in a sandbox and opens a merge request. If the tests fail twice, the task is either handed to a human or, where policy allows, routed to the frontier tier with the code redacted to the relevant files. The team reviews pass rates and reviewer acceptance monthly, and keeps the local tier only if it holds up against that record.

That is also where we would start with a client. If your team is weighing a self-hosted coding agent, or wants an evaluation of Ripple-65B Preview against its own repositories before committing hardware, talk to us about a scoped test. For background on how we approach agent tooling for engineering teams, see our developer experience page.

KEY FACTS
Model
Ripple-65B Preview is a 65B-parameter dense Transformer (not Mixture-of-Experts) with d_model 5120 and 80 layers, according to Ricursive Intelligence.
Licence and context
The weights are released under Apache 2.0 on Hugging Face, and the model card states support for up to 1M tokens of context.
Headline score (self-reported)
Ricursive reports 80.52 on Terminal-Bench 2.1, a U.S. open-weight best, against 80.1 for Reflection Beam 501B and 76.4 for Muse Spark Open 816B on its own chart. The model card says agentic scores have been hard to replicate.
Hardware
Ricursive says the model fits on consumer hardware such as the NVIDIA DGX Spark. No throughput figure on that device was published; it reports roughly 250 tokens per second on two B200 GPUs before quantisation.
Preview limits
The Preview has had relatively light post-training focused on agentic coding, and was pretrained on a predominantly English and code corpus.
Origin and cost
Ricursive describes it as a general-purpose variant of the proprietary model behind its chip-design platform, says it bought no post-training data, and puts development and training cost at $32M (unaudited).
Questions

Questions about this project.

What is Ripple-65B Preview?

It is a 65-billion-parameter dense Transformer released by Ricursive Intelligence under the Apache 2.0 licence, with weights on Hugging Face. Ricursive says it supports up to 1M tokens of context and has been lightly post-trained with a focus on agentic coding. It is a general-purpose variant of the proprietary model behind the company's chip-design platform.

Is Ripple-65B the best open-weight coding model?

Ricursive claims a U.S. open-weight best of 80.52 on Terminal-Bench 2.1, which is a narrower claim. It excludes Chinese open-weight families such as Qwen, DeepSeek, GLM and Kimi. The score is self-reported, the lead over the next model on Ricursive's chart is under half a point, and the model card says agentic benchmark scores have been hard to replicate.

Can Ripple-65B Preview run on-premise?

Ricursive says it fits on consumer hardware such as the NVIDIA DGX Spark, and reports roughly 250 tokens per second on two B200 GPUs before quantisation. It has not published throughput on the DGX Spark, so a team should measure speed on its own hardware before sizing a deployment. As a dense model it is simpler to serve than a large Mixture-of-Experts model.

What should a team test before using it?

Run it against tasks from your own repositories in a sandbox, include tickets and documentation written in Spanish or Catalan because the corpus is mostly English and code, test failure handling, destructive commands and prompt injection because post-training was light, and review the model's provenance and pin the weights to a checked version.

Continue reading

Shipped & in build

Real work. Check it yourself.

What we built, what stage it is at, and a link to the live thing wherever there is one.

The updated yourOS homepage: Roxy’s open ring surrounded by bento cards for calls, voice, translation, building, design, apps and more.
The yourOS homepage on a phone, with its compact navigation, welcome film button and bento cards.
youros.tech — the homepage bento hero
Our own · coming soon

yourOS

yourOS is an operating system with an AI at its heart. You talk to it like a person, out loud or typed. It makes your calls, plans your days and works on your projects with a crew of agents, on your own private server. Our own product, built from the first commit; the waitlist is open.

What we built
  • Operating system
  • Roxy
  • Agent crew
  • Voice calls
  • App store
  • Website
Meet yourOS →
The Sonora homepage on desktop: a full-bleed dune landscape behind the headline “Transform Your Life with Sound”, with App Store and Google Play download buttons.
The Sonora homepage on a mobile viewport, showing the same hero reflowed to a single column.
sonora.com — homepage, 1440×900
Live · iOS · Android

Sonora

Website, backend and app assets for Sonora, a free AI sound-therapy app on iOS and Android published by Cognitive AI.

What we built
  • Website
  • Backend
  • App assets
Read the case study →
The updated Podswap homepage on desktop: “Grow on Instagram and TikTok for free, with real engagement from creators your size”, beside a tilted wall of exchange, tool and guide cards on a violet gradient.
The updated Podswap homepage on a phone, with its floating navigation, Instagram and TikTok engagement headline and tilted tool cards.
podswap.app — homepage, 1440×900
Five platforms

Podswap

Brand, website, prototype, apps on five platforms, the AI backend and the algorithm-intelligence pipeline for Podswap, a social growth engine that runs without a human in the loop.

What we built
  • Brand
  • Website
  • Prototype
  • Apps on five platforms
  • AI backend
  • Algorithm pipeline
Read the case study →
The unOpen.AI homepage introducing Roxy, a private AI assistant that remembers who you are.
The unOpen.AI homepage on a mobile viewport.
unopen.ai — homepage, 1440×900
Prototype · closed beta

unOpen.AI

In closed beta: Roxy, a private AI assistant with permanent memory running on infrastructure the user owns. Brand, site, product, backend and conversational design.

What we built
  • Brand
  • Site
  • Product
  • Backend
  • Conversational design
Read the case study →
coherence.com on desktop, headed "Intelligence that moves in phase with people".
coherence.com on a mobile viewport, reflowed to a single column.
coherence.com — homepage, 1440×900
Live

Coherence

Live at coherence.com: brand identity generated with AI agents we directed, the motion system and website for the group behind Sonora, AboutAI, WakeUp and MudShadows. Agent work remains in build.

What we built
  • Brand identity
  • Motion system
  • Live website
  • Agent work — in build
Read the case study →
The new Domain Booth homepage on desktop: the headline “Premium Domains. Serious Buyers. Proven Results.” beside the animated DB monogram.
The new Domain Booth homepage on a phone, reflowed to a single column.
Our build of the Domain Booth homepage — 1440×900
Live

Domain Booth

Live at domainbooth.com for DomainBooth, a premium domain brokerage: a new animated mark, every page redesigned and rebuilt with designed mobile layouts, and a custom CMS for its sales record, featured names, press and enquiries.

What we built
  • Animated mark
  • Every page redesigned
  • Designed mobile layouts
  • Custom CMS
Read the case study →
Who we have built for

Twenty-one years. Names you know.

See all of our work →
  • BBC Good Food logo
  • Charities Aid Foundation logo
  • GAME logo
  • NatWest logo
  • Next logo
  • NSPCC logo
  • Tesco logo
  • Worldpay logo
  • Samsung logo
  • Sony logo
  • Land Rover logo
  • UAP Limited logo
  • Netflix logo
  • Amazon Prime Video logo
  • Mercedes-Benz dealerships logo
  • San Miguel logo
  • Biovene Barcelona logo
  • Seventy7 logo
  • Littlewoods logo
  • Smile logo
  • The Co-operative Bank logo
Let’s talk

Working on
something like this?

If this is the kind of problem you are working on, we are happy to talk it through.

30-minute technical call · no deck