Open-weight models bring Artificial Intelligence (AI) to personal computers and self-managed data centers. New releases such as Inkling 2 Small, Ripple-65B Preview, and NVIDIA Dory 4B also show that “open” does not automatically mean local, inexpensive, or reliable. For you, the relevant issue is therefore not the leaderboard alone, but whether a model fits your device, task, and need for control.
Open-weight and local AI explained
An open-weight model makes its trained model weights available, meaning the numerical values that store its learned behavior. That is not necessarily the same as fully open-source software: licenses, training data, tools, and usage rights may remain restricted or incompletely documented. A local model runs on your own computer or infrastructure you control instead of sending every input to an external cloud service.
The terms consequently describe two separate qualities. A model can have open weights yet be so large that it remains impractical on typical personal hardware. Conversely, a smaller model may run locally even if its provider has not opened every part of the system.
This distinction is particularly clear with Thinking Machines’ Inkling 2 Small. The model has 594 billion parameters, the adjustable values within a model, but activates only 17 billion for each processing step. This Mixture-of-Experts (MoE) architecture directs work to selected parts of the model in an attempt to use computing resources more selectively.
At launch, however, Inkling 2 Small was available only through the Tinker platform for use and fine-tuning. Fine-tuning means additional training that adapts an existing model to particular tasks or behaviors. An Apache 2.0 weight release was planned after external safety testing, while the hardware required for self-hosting had not been disclosed, according to the technical overview of Inkling 2 Small.
Performance between flagship and specialist tool
Inkling 2 Small accepts text and images and offers variable reasoning effort, allowing users to trade speed for more thorough processing. It scored 31 on the Artificial Analysis Intelligence Index according to the provider, up from 13 for its predecessor. That is roughly comparable to Google’s Gemini 3.8 Flash, but it is not an independent guarantee of equally good results in your daily work.
Thinking Machines reports major gains on agentic tasks, meaning multistep workflows performed partly autonomously by a model. The reported AutomationBench-AA result increased from 5 to 46 percent, while FrontierCode 2.0 rose from 5 to 25 percent. Performance nevertheless varies sharply by test: the table reproduced by MarkTechPost gives the model 86.3 percent on SWE-Bench Pro v2 Hard but only 13.1 percent on Terminal-Bench 4.0.
Ripple-65B Preview takes a different route. It has 65 billion parameters, is optimized for NVIDIA hardware, and is available on Hugging Face under an Apache 2.0 license, according to the report on Ripple-65B Preview. It trails Inkling 2 Small in several listed evaluations but is expected to be easier to set up on privately managed hardware.
For a workplace example, a team might use Ripple on its own infrastructure to process and review coding tasks rather than submit source code to a general cloud service. Inkling would be more attractive when stronger agentic performance matters more than immediate self-hosting. Both examples concern specialized workflows and do not automatically translate into better everyday email writing or document summaries.
NVIDIA Dory 1.0 4B is smaller still. It can run the same internal block one to four times per token, providing four computing budgets without retraining the model. A guide to running Dory locally uses a 3.7-gigabyte GGUF file, a model file format optimized for local execution, and examines how extra passes affect speed, answer length, and accuracy.
Pros and Cons of open and local models
Pros:
- Greater control – You can choose the model version, infrastructure, and update schedule when its weights and license permit self-hosting.
- Specialization – Fine-tuning can adapt a base model to a clearly defined field or set of response rules.
- Flexible computing cost – Models such as Dory provide different compute budgets, while MoE models activate only part of their parameters at once.
- Broader choice – Specialized models are available for coding, robotics, and driving simulation alongside general language models.
Cons:
- High resource demands – Open weights are of limited help when memory, suitable hardware, or practical runtime software is missing.
- Token appetite – More generated text units can make tasks slower and more expensive without producing a proportionally better answer.
- Uneven reliability – Strong overall scores can conceal weak results on particular tests and unusable outliers.
- Demanding customization – Fine-tuning requires suitable data, explicit evaluation standards, and checks after training.
The cost trade-off is particularly visible with Inkling 2 Small. According to the lead report, it generates more output tokens, the text units that make up a response, and costs six times as much per task as its predecessor. A better score may therefore be economically unattractive if you process many simple jobs rather than a smaller number of demanding workflows.
Fine-tuning as an opportunity and source of influence
Open weights support not only professional customization but also changes to political and cultural response patterns. Nous Research gave Alibaba’s Qwen3.8-27B an additional 53 million tokens of training and released the result as Arete-27B. According to the report on this fine-tuning project, the company’s measured propaganda rate fell from 22 percent to less than 1 percent without reducing general capabilities.
Those numbers are based on Nous Research’s own evaluation method. Its test included 803 prompts about topics such as freedom of expression, education, privacy, surveillance, and democracy; another AI model judged the answers. The example therefore demonstrates both the power and the sensitive side of fine-tuning: behavior can be changed, but someone must decide which data, facts, and values provide the standard.
Specialization also reaches far beyond text models. NVIDIA released four open checkpoints in its Cosmos 3 family for interactive world simulations, including models for varied weather and lighting, synchronized camera and LiDAR data, and predictions of the visual effects of robot actions. According to NVIDIA’s introduction to the Cosmos models, each variant has 16 billion trainable parameters and permits commercial and noncommercial use under the terms of OpenMDW1.1.
One concrete professional example is generating multiple views of a driving scene under different weather, lighting, and time-of-day conditions. Another is a robotics simulation in which the model predicts how an action will change what a camera sees. These models are not universal chat assistants, but they may be more useful than a larger general language model for a narrowly defined job.
Reliability and a practical way to begin
Strong benchmark scores do not prevent strange individual failures. An analysis of 2,655 Inkling runs across 13 benchmarks found English profanity or variants in 50 runs; 241 of the 316 total instances came from just two runs that went off the rails. During an asteroid calculation, the model produced a response longer than 500,000 characters and repeatedly recycled the same sentences.
The profanity was nevertheless not a useful warning signal, according to the analysis of Inkling’s reasoning logs. Although 46 of the 48 evaluated runs containing profanity failed, about 75 percent of all evaluated runs failed, and 96 percent of failed runs contained no profanity at all. Users generally do not see these internal reasoning steps, or see them only through a filter.
For beginners: The most sensible first step is a small model assigned to a tightly limited task. You could test Dory locally at different compute levels while repeating the same basic jobs, such as summarizing short passages or checking specified calculation steps. The key measure is not merely whether an answer sounds plausible, but how often it is correct, fast, and reproducible.
For advanced users: You can gain more value by building your own evaluation set from typical work cases and difficult edge cases. This lets you compare model size, response length, runtime, and error rate before considering fine-tuning. For agentic workflows, intermediate results and external actions should be monitored because one model error can otherwise propagate through several later steps.
In Switzerland, local models may appeal to organizations that want to keep inputs and model operations under greater internal control. The sources do not provide specific Swiss availability information or verified details about support for German, French, Italian, or Romansh. Local operation may limit data transfers, but it does not replace a review of licenses, access permissions, stored logs, and applicable privacy requirements.
Remaining questions about selection and cost
Direct comparisons are complicated by different architectures, licenses, and evaluations. Inkling combines a large total size with relatively few active parameters, Ripple uses a dense 65-billion-parameter model, and Dory trades additional computing time for potentially better output through repeated loops. A single benchmark captures these differences only partially.
The term “inexpensive” also needs a specific point of reference. A small model file may be affordable to run locally while setup and evaluation still consume time. A more capable model might solve a difficult task more successfully but cost substantially more per job because of long responses or heavier computation.
Inkling also illustrates the gap between an open-weight announcement and immediate availability. Until hardware requirements and practical self-hosting results are known, its suitability for local operation cannot be assessed reliably. Ripple and Dory are more tangible self-hosting examples, but they address different levels of performance.
Open-weight AI expands your choices between cloud convenience, local control, and professional specialization. Current examples range from a compact local reasoning model to interactive driving simulation, yet openness alone resolves neither hardware requirements nor reliability problems. The largest unresolved risk is that convincing average scores may hide rare, expensive, or difficult-to-detect failures.
Sources
- Inkling 2 Small soll das stärkste westliche Open-Weight-Modell sein und erreicht Googles Gemini 3.8 Flash – The Decoder, 2026-11-06
- Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active – MarkTechPost, 2026-11-06
- “FUCK. NOTHING WORKS.” – KI-Modell Inkling 2 Small flucht in seinen Denkprotokollen – The Decoder, 2026-11-07
- Neues Open-Weight-Modell Ripple-65B Preview soll besonders gut agentisch programmieren – The Decoder, 2026-11-06
- How to Run NVIDIA Dory 4B with llama.cpp: Loop by Loop Guide – MarkTechPost, 2026-11-07
- Feintuning gegen Propaganda: US-Start-up trainiert Alibabas Qwen-Modell auf “westliche Werte” um – The Decoder, 2026-11-07
- New NVIDIA Cosmos 3 Open Models for Interactive World Simulation – Hugging Face, 2026-11-05
Image: Matheus Bertelli via Pexels