Within a few days, three open-weight language models aimed at development teams have appeared.Open weights means: the trained model parameters are available for download and can be run on your own infrastructure instead of exclusively via a provider's API.For projects with strict data protection requirements or predictable cost targets, this is the decisive difference.
What the three releases have in common is their application focus: agentic programming.This refers to models that not only suggest code snippets but work through multi-step tasks – reading files, executing commands in a shell, evaluating test results and deriving the next step from them.
Ripple-65B Preview: lean and trimmed for terminal tasks
Ricursive Intelligence has released Ripple-65B Preview, an open model with 65 billion parameters.According to the company, it achieves the best score among open US models on the TerminalBench 2.1 coding benchmark.TerminalBench measures how reliably a model actually completes tasks in a command-line environment – a scenario that comes closer to everyday work in development projects than pure code completion.
Notable is the stated cost figure: 32 million dollars in total costs for the model.On other benchmarks, however, Ripple lags significantly behind the simultaneously released Inkling 2 Small.The model is therefore not an all-rounder but a specialist – which can certainly be sufficient for clearly defined use cases.
Inkling 2 Small: strongest Western option, but hungry
Thinking Machines, the start-up of former OpenAI CTO Mira Murati, positions Inkling 2 Small as the strongest Western open-weight model.Technically, it is a mixture-of-experts model with 594 billion parameters.In this architecture, the model is divided into many subnetworks – the experts – of which only a fraction is activated per processing step.This reduces the computational effort compared with an equally large dense model.
However, the better agent performance is bought at the price of significantly higher token consumption.This is a point that rarely appears in benchmark tables but directly hits the bill in practice: a model that produces more intermediate steps and longer chains of thought for the same task consumes more computing time – regardless of whether billing is via API or self-hosting.
The name "Small" evidently refers to its classification within its own model family, not to the absolute resource requirement.Running a model of this magnitude yourself requires corresponding GPU infrastructure.
ERNIE 5.5 Flash Base: raw material instead of a finished assistant
Baidu takes a different approach with ERNIE 5.5 Flash Base.The mixture-of-experts model has 280 billion parameters, of which only 14 billion are active per token – a ratio designed for efficient operation.It was pretrained on 33.5 trillion tokens.
The decisive factor is its release as a base model without fine-tuning.Such a model is not a ready-to-use chat or coding assistant but starting material: it has absorbed language and world knowledge but was not trained to follow instructions or respond in a particular style.Baidu leaves this adaptation to the community and thus to organisations that want to contribute their own training data and objectives.
For companies, this means two things.The effort required to reach productive use is higher than with a fully instructed model.On the other hand, there is greater scope to align the behaviour with a specific domain, a particular codebase or internal conventions.
How the figures relate to one another
The three reports come from the same source and are close together in time, but they are not directly comparable.Ripple-65B is classified via a specific terminal benchmark, Inkling 2 Small via a broader performance claim, ERNIE 5.5 Flash Base primarily via architecture and training scope.A clean three-way comparison cannot be derived from this.
In addition: all performance figures come from the respective providers.Independent verification is usually pending, and benchmark best scores in a narrowly defined category – such as "open US models" – say little about behaviour in a concrete project.
What matters in the assessment
For selecting an open model, other questions are relevant in practice than pure benchmark placement:
- Actual resource requirements in operation: Parameter count alone is not enough.For mixture-of-experts models, all parameters must be held in memory, even though only a subset computes per token.
- Token consumption per solved task: A model with better results that requires significantly more steps can end up being more expensive in the overall calculation.
- Maturity: Preview versions and base models without fine-tuning require additional work before they can be used productively.
- Licence and provenance questions: Open weights are not automatically compatible with every commercial use; the specific conditions must be checked.
- Suitability for your own technology stack: A model that is strong at terminal tasks does not necessarily handle the framework used in the project equally well.
Assessment for web and software projects
The density of releases shows that open models in the coding space are increasingly a serious alternative to proprietary APIs — especially where source code or customer data must not leave your own network. Realistically, however, this is not a sure thing: operation, customisation and ongoing evaluation of model quality tie up resources that would lie with the provider when using an API.
For most mid-sized projects, a two-stage approach is therefore advisable: first validate the specific use case with an available service, while designing the architecture so that the model remains interchangeable. This allows a later switch to a self-hosted open model without disruption if data protection, costs or availability suggest it.
Sources
- Neues Open-Weight-Modell Ripple-65B Preview soll besonders gut agentisch programmieren — The Decoder
- Inkling 2 Small soll das stärkste westliche Open-Weight-Modell sein und erreicht Googles Gemini 3.8 Flash — The Decoder
- Baidu veroeffentlicht ERNIE 5.5 Flash Base mit 280 Milliarden Parametern als Open-Source-Basismodell — The Decoder