What is Ripple-65B Preview?
Ricursive Intelligence published its launch post for Ripple-65B Preview on 6 November 2026. The weights had already appeared in a Hugging Face repository created on 4 November, and OpenRouter gives the same release date. Ricursive describes the model as "a general-purpose variant of the proprietary model that powers our AI chip design platform".
Ripple-65B Preview is a dense Transformer with 64,782,689,360 parameters, according to the safetensors metadata on Hugging Face. The company rounds that to 65 billion. It uses 80 layers and a model width of 5,120, a deliberately narrow and deep shape designed for efficient training.
Ricursive pretrained the model from scratch on 12 trillion tokens, mostly English text and code, at an 8,000-token sequence length. The released model accepts up to 1,048,576 tokens of context. That figure describes the maximum the architecture supports, not a promise that every hosted endpoint will serve it.
The release includes three checkpoints. One is an instruction-tuned model in BF16, one is a compressed NVFP4 version, and one is a base model that skipped midtraining and synthetic question-answer data. All three sit under the Apache 2.0 licence, which allows commercial use, modification and redistribution without user-count limits.
That licence is part of the pitch. Several open-weight releases use community licences with restrictions, while Apache 2.0 gives companies a clearer path to internal deployment. Ricursive lists Hugging Face, Together AI, GMI Cloud, OpenRouter and Fireworks as distribution channels. VentureBeat found a public rate only for the OpenRouter endpoint, which GMICloud serves.
Despite the open licence, the name still says Preview. Ricursive describes post-training as relatively light and focused on agentic coding. Buyers should treat the release as an early checkpoint rather than a finished general-purpose assistant.
How does Ripple-65B score on coding benchmarks?
Ricursive's headline claim is a score of 80.52 on Terminal-Bench 2.1, a benchmark that asks an agent to complete tasks inside a terminal. The company calls the result "U.S. open-weight state-of-the-art" and compares it with four larger American models.
These tests measure more than autocomplete. Terminal-Bench 2.1 asks an agent to edit files, install packages and run commands until a task is finished. SWE-bench Pro v2.0 Hard uses repository-level fixes that require several steps. Both reward tool use, recovery from errors and long-horizon planning, which is why labs now cite them next to older code-generation scores.
The launch post puts Ripple at 80.52, ahead of Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8 and NVIDIA Nemotron 3 Ultra 550B at 56.4.
The margin over Reflection's Beam is 0.42 points on those rounded figures. That is close enough that the choice of test harness matters, and the model card shows the gap moving when the harness changes.
With the official Terminus 2 harness, Ripple still scores 80.5. Meta's Muse Spark rises to 78.7, Beam falls to 70.8, Thinking Machines' Inkling scores 56.2 and Nemotron scores 55.1. The lead narrows to 1.8 points. Ricursive says it used Terminus 2 for SWE-bench Pro v2.0 Hard and DeepSWE 1.1, and OpenCode elsewhere when a harness was not specified.
Ripple also leads the card's SWE-bench Pro v2.0 Hard column with 79.1. Beam scores 76.5 and Muse Spark 72.4. That benchmark targets difficult software engineering tasks and was introduced in September. A strong result there is more relevant to maintenance work than a short terminal exercise.
Other rows tell a different story. Ripple scores 76.5 on SWE-bench Verified, behind Beam at 81.6, Muse Spark at 78.7 and Inkling at 77.6. It scores 61.42 on SWE-bench Pro v1.0, behind Beam at 66.1. On DeepSWE 1.1 it scores 44.3 against Muse Spark's 60.0 and Beam's 43.3.
The Decoder noted that Ripple trails the newly released Inkling 2 Small on other benchmarks, while likely being easier to deploy on local hardware. That trade-off fits the pitch. Ricursive is selling a small model that a team can host, not a model that wins every table.
The table mixes sources. Ricursive says it uses official model-card scores where available and its own evaluations otherwise. Italic figures in the card mark in-house runs; plain figures come from rivals' published results. Ripple's own scores come from its harness choices, so the comparison is not a controlled experiment.
No independent evaluator had published Ripple results when this article was written. Artificial Analysis, which runs a common third-party comparison, did not list the model. Until that changes, every score in this section is either Ricursive's own measurement or a rival's measurement quoted by Ricursive.
Why does a dense 65B model matter?
Ricursive chose a dense Transformer rather than a mixture-of-experts design. Every parameter takes part in every token. That simplifies training and serving, but it raises compute per token compared with a sparse model that activates only part of its weights.
Total size and active size separate Ripple from its rivals. Ripple has about 65 billion active parameters. Reflection's Beam has 501 billion total parameters but only 23 billion active at a time. Beam can be cheaper per generated token, while Ripple needs a smaller memory footprint and a simpler serving stack.
The BF16 checkpoint is about 130 GB, based on the file sizes Hugging Face lists. That does not fit on a typical laptop. Ricursive offers an NVFP4 version and says the model can run on an NVIDIA DGX Spark. It also mentions MacBooks in social posts, but the launch material does not pin down a single laptop memory configuration.
Context length adds another constraint. Serving the full 1,048,576-token window requires far more memory than the weights alone, because the key-value cache grows with the sequence. Ricursive has not published a full memory table for that setting. OpenRouter lists the 1M context and a 131,072-token output limit, while the model card reports 250 tokens per second on two B200 GPUs at batch size eight before quantization.
Ricursive says training held about 40% model FLOPs utilisation without custom kernels or quantized training. The team grew from one to five people during development, according to the model card. Those details support the efficiency story, but they describe training, not the cost of serving many users.
The base checkpoint has its own comparisons. Ricursive says it beats Poolside's Laguna M.1 and is comparable with DeepSeek V3.1 and V4 Flash on reasoning, maths and code. It also says NVIDIA's Nemotron 3 Ultra leads traditional pretraining evaluations, likely because of heavy synthetic data. Those results matter to teams that fine-tune a base model, but they do not transfer directly to agentic coding.
Hosted capacity is thin for now. OpenRouter lists a single provider, GMICloud, and 78.5% uptime over three days at the time of writing. A buyer who needs steady production throughput should test the endpoint before depending on it.
How much did Ripple-65B cost to train?
Ricursive puts the full research and development cost at $32 million. That covers pretraining experiments, data mixing, evaluations, debugging, synthetic data, post-training recipe work and the final training runs, according to the launch post. It is an all-in figure, not just the final compute bill.
The comparison is stark. Ricursive cites Sam Altman's 2023 remark that GPT-4 training exceeded $100 million and Dario Amodei's warning that frontier training could reach $1 billion or more. It then estimates that Ripple cost about one-tenth of the previous open-weight leader on agentic coding.
That estimate is not an audited comparison. Ricursive bases it on reported compute commitments and training durations for Reflection's Beam. VentureBeat notes that Reflection reportedly committed $150 million per month for compute at SpaceX's Colossus cluster and trained Beam for four weeks of pretraining and four weeks of post-training. Those figures describe different accounting from Ricursive's all-in research number.
The reinforcement-learning counts are also hard to compare. Ricursive says it reached its Terminal-Bench score with 1.4 million rollouts, roughly 70 times fewer than Beam. Reflection has reported more than 100 million rollouts during a four-week run. Rollouts differ in length, hardware demand and evaluation cost, so the counts do not establish an equivalent saving.
Ricursive also says it bought no post-training data. It built its own synthetic pipelines, quality filters and coverage metrics, then tuned reinforcement learning with truncated importance sampling, score centring and a length penalty. Those choices lower data costs, but they make the recipe harder for outsiders to reproduce without the same tooling.
Ricursive reports a 16-fold improvement in compute efficiency on code at the 4-billion-parameter scale, compared with an Olmo 3 baseline. The model card says the advantage grows to 55-fold for the full training run. That measures loss on a training objective, not coding-agent performance or the price of an API call.
The cost claim still matters. A $32 million run is small enough for a venture-backed startup to fund without a hyperscaler's balance sheet. If the recipe generalises, more labs could train competitive coding models at a fraction of frontier budgets. If it depends on Ricursive's private data and chip-design workload, the result will be harder to copy.
Who is Ricursive Intelligence?
Ricursive is not a typical model lab. Its founders are Anna Goldie, the chief executive, and Azalia Mirhoseini, the chief technology officer. Both worked on AlphaChip, Google's reinforcement-learning system for chip placement, before starting the company.
VentureBeat reports that Ricursive raised a $300 million Series A in January 2026, led by Lightspeed at a $4 billion valuation. Nvidia's NVentures, DST Global and Sequoia also participated. The company's main product is a platform for designing chips, not a chatbot.
Ripple is therefore a public byproduct of internal work. Ricursive says an internal variant powers the chip-design platform. The coding benchmarks measure the general-purpose model, not the platform's ability to design hardware. Buyers should treat those two claims separately.
The company is also using the release to recruit. Its launch post ends with a hiring note and a link to open roles. A strong open model is a visible calling card for a lab that needs researchers who can work across silicon and software.
What are the limits of Ricursive's claims?
The phrase "US open-weight state-of-the-art" carries two limits. It covers American models only, and it covers the benchmarks Ricursive chose. Reflection's published comparison lists several Chinese-developed models above Ripple's Terminal-Bench figure, according to VentureBeat. A global open-weight ranking would look different.
Benchmark harnesses add another layer. Agentic coding tests give a model tools, a file system and a time limit. Small changes to the harness can move scores by several points, as the Terminus 2 row shows. Ricursive acknowledges the problem in its model card. "Replicating benchmark scores has been challenging," it says, citing differences in evaluation setup.
Some numbers have also moved between Ricursive's materials. VentureBeat found that the model card it reviewed listed different rival scores from the launch graphic, including for Meta's Muse Spark and Thinking Machines' Inkling. The current card lists 72.4 for Muse Spark and 39.6 for Inkling on SWE-bench Pro v2.0 Hard, matching the graphic. Changing tables make exact cross-model comparisons fragile.
The preview label matters too. Ricursive describes post-training as relatively light and focused on agentic coding. The model card concentrates on capabilities and does not publish a red-team or safety evaluation. Apache 2.0 gives users broad freedom, but it does not certify quality or guard against misuse.
The licence also shifts responsibility to the user. Weights can be redistributed and fine-tuned without a licence review, which suits enterprises but leaves them without a vendor safety net. Apache 2.0 itself has no usage cap, no acceptable-use addendum and no requirement to report evaluation results.
Finally, a coding benchmark is not a workplace. Real agents depend on the surrounding software, the tools they can reach, the length of the task and how failures are handled. A strong Terminal-Bench score is evidence worth testing, not proof of lower cost or fewer bugs.
The hosted service is an early signal of the same gap. One provider and uneven uptime are normal for a launch week, but they matter for anyone planning production traffic. Self-hosting avoids that dependency and shifts the burden to hardware, memory and operations.
What to watch next for open-weight AI?
The next signal will come from independent evaluation. Artificial Analysis, LMArena and third-party coding harnesses can test the model without Ricursive's choices. Until those results appear, the 80.52 figure should be read as a company claim with a published method.
Competition is already close. Thinking Machines released Inkling 2 Small the same week, and Reflection and Meta have larger open-weight models with strong coding results. If Ripple's efficiency claims hold, the interesting question is whether other labs copy the dense, small-team approach or stay with mixtures of experts.
Enterprise buyers should watch hosted reliability and price. A $0.20 input rate is low, but a single provider with uneven uptime changes the maths for production workloads. Local deployment on workstations is the other path, and it depends on quantized checkpoints and memory budgets that Ricursive has not fully documented.
Documentation is the other gap. A production release would need memory tables for long context, per-provider context limits and reproducible evaluation scripts. Ricursive has published a model card, a Colab and an NVFP4 checkpoint, but not a full deployment guide. Those artefacts are less visible than a benchmark chart and more useful to buyers.
Ricursive has framed Ripple as a step toward models that help design chips. The company's next release, and any independent test of its chip-design platform, will matter more than another benchmark table. For now, Ripple-65B Preview is a credible open-weight entrant with a narrow lead, a low price and a short track record.