Longcat 2.5, Ripple 65B, Inkling Small: This Week's Open-Weight Model Drops
Longcat 2.5, Ripple 65B, and Thinking Machines' Inkling Small landed this week. Here's what each open-weight model offers and who can run them.
What are Longcat 2.5, Ripple 65B, and Inkling Small?
They’re three open-weight language models released within days of each other, each aimed at a different slice of the market. Longcat 2.5 is a massive 1.6 trillion parameter multimodal model from a relatively unknown lab called Longcat. Ripple 65B Preview is a 65 billion parameter dense model from a US company, Recursive Intelligence, built for strong coding and agentic performance at a small footprint. Inkling Small comes from Thinking Machines, Mira Murati’s company, and targets the small-model tier where it goes head to head with DeepSeek’s open-weight models. None of these require a subscription or API key from a closed lab. You can download the weights and run them yourself, assuming you have the hardware.
TL;DR
- Longcat 2.5 is a 1.6 trillion parameter multimodal model with 48 billion active parameters, strong on visual understanding, creative design, and 3D tasks, but too large for almost anyone to run locally.
- Ripple 65B Preview from Recursive Intelligence is a 65 billion parameter dense model scoring 80.5% on Terminal Bench 2.1 among open-weight models, with a million-token context window and pricing around 20 cents per million input tokens and 60 cents per million output tokens.
- Ripple is explicitly built for local feasibility, with the company saying it can run on a MacBook or a DGX Spark, making it one of the more practical high-capability open-weight options right now.
- Inkling Small from Thinking Machines scores 31 on the Artificial Analysis Intelligence Index and competes closely with DeepSeek V4.1 Flash, currently considered the strongest open-weight model overall, despite being a much smaller model.
- DeepSeek V4.1 Flash remains the benchmark to beat among open-weight models, and Inkling Small’s ability to approach its scores on Frontier Code, SciCode-style benchmarks, and agent tasks is notable given the size gap.
- The active-parameters-vs-total-parameters split matters: Longcat’s 1.6 trillion total parameters sound enormous, but only 48 billion are active per forward pass, which is the more relevant number for inference cost and speed.
- This cluster of releases points to a broader trend: multiple US and international labs outside the usual frontier names (OpenAI, Anthropic, Google) are now shipping competitive small-to-medium open-weight models on a near-weekly cadence.
● Reading src/server.ts
● Editing 3 files
+ import { limit } from 'rate'
- app.use(cors())
✓ Tests passed · ready for PR
How big is Longcat 2.5, and can you actually run it?
Longcat 2.5 is a 1.6 trillion parameter model, which places it in the same size class as some of the largest models ever released publicly. It uses a mixture-of-experts style architecture, with only 48 billion parameters active at inference time. That active parameter count is what actually determines how expensive and slow the model is to run, not the total parameter count, but 1.6 trillion total parameters still means an enormous amount of storage and memory is needed just to hold the weights, even quantized.
Longcat 2.5 is multimodal, with stronger visual understanding than typical text-only open models, and the team behind it emphasizes creative design and 3D task performance. For most individual developers and hobbyists, this model is not something you load onto a local machine. It’s realistically a model you’d run through a hosted inference provider or on serious multi-GPU server infrastructure. It matters as a signal that open-weight labs are willing to ship frontier-scale parameter counts, but local feasibility is essentially zero for a typical user.
Is Ripple 65B worth running locally?
Ripple 65B Preview is the most locally-practical release of the three. At 65 billion dense parameters, it sits in a size range that’s been runnable on high-end consumer and prosumer hardware for a while now, especially with quantization. Recursive Intelligence, the company behind it, specifically says the model can fit on a MacBook or a DGX Spark, which signals they designed it with local deployment in mind rather than purely as a cloud-hosted product.
On benchmarks, Ripple scored 80.5% on Terminal Bench 2.1, measured against other open-weight models rather than closed frontier models. Terminal Bench tests a model’s ability to complete real command-line and coding tasks, so a strong score there suggests Ripple is tuned for agentic and developer workflows rather than general chat. It also supports up to a million tokens of context, which is unusually long for a model this size and useful for tasks like codebase-wide analysis or long document review.
Pricing, for anyone using it through an API rather than running it locally, comes in at 20 cents per million input tokens and 60 cents per million output tokens, which is inexpensive relative to most frontier closed models. Recursive Intelligence isn’t a widely known name yet, but a US-based team shipping a genuinely strong, genuinely small open-weight model with real benchmarks attached is a meaningful data point in the open-weight space.
How does Inkling Small compare to DeepSeek’s open models?
| Vendor | Tier | Risk | ApprovalNEW |
|---|---|---|---|
| Northwind Supply | Tier 1 | Low | Approved |
| Acme Logistics | Tier 2 | Medium | Approve |
| Helios Data | Tier 1 | High | Pending |
| Granite Freight | Tier 3 | Low | Approved |
| Pinecrest Labs | Tier 2 | Low | Approve |
Inkling Small, released by Thinking Machines, scores 31 on the Artificial Analysis Intelligence Index, a composite benchmark that aggregates performance across multiple tasks. On its own that number means less without context, but the more useful comparison is against DeepSeek V4.1 Flash, which is currently regarded as the strongest open-weight model available. Despite being considerably smaller, Inkling Small holds up reasonably well against it: on Frontier Code it scores 25 versus DeepSeek’s 28, and on a SciCode-style benchmark it hits 72 versus DeepSeek’s 77. It also posts 86 on SWE-Bench Pro and 31 on a post-training benchmark.
Those gaps, roughly 3 to 5 points across coding and agentic benchmarks, are small for the size difference involved. It suggests Thinking Machines has done meaningful work on training efficiency rather than just scaling parameters. Demonstrated use cases for the model include building small games and simulations, a from-scratch PC emulator written in Python, and visually distinct creative projects, all generated directly by the model. For developers who want a capable, small, open-weight model that’s realistically self-hostable, Inkling Small is one of the more interesting recent options.
Why are so many small open-weight models showing up at once?
Part of the answer is economic. Running a frontier-scale closed model like Anthropic’s Claude Opus or Google’s Gemini line for every single task is expensive, especially at high reasoning effort settings. A common workflow emerging among heavy AI users is to plan or architect a task with a frontier model, then delegate the bulk execution to a cheaper, smaller model. That only works if the smaller model is actually competent, which is exactly the lane Ripple 65B, Inkling Small, and similar efficient models are built for.
There’s also a geographic and competitive angle. Longcat appears to be a newer, less established lab pushing out a frontier-scale parameter count model to establish credibility. Recursive Intelligence and Thinking Machines are both US-based teams explicitly building smaller, efficient, open-weight models rather than chasing the largest possible parameter count. That’s a different strategy than the “biggest model wins” approach, and it’s a sign the open-weight ecosystem is starting to specialize: some labs chase frontier scale, others chase efficiency and local runnability.
Frequently Asked Questions
Can I run Longcat 2.5 on a personal computer?
No. At 1.6 trillion total parameters, Longcat 2.5 requires server-grade infrastructure or a hosted inference provider. Even with 48 billion active parameters per forward pass, the full weight set is too large for consumer hardware.
What makes Ripple 65B different from other open-weight models this size?
Ripple 65B Preview pairs a relatively small 65 billion dense parameter count with a million-token context window and strong Terminal Bench 2.1 performance (80.5%) among open-weight models, while being explicitly designed to run on hardware like a MacBook or DGX Spark.
Is Inkling Small better than DeepSeek V4.1 Flash?
Not outright. DeepSeek V4.1 Flash still leads on most benchmarks mentioned, including Frontier Code (28 vs 25) and a SciCode-style test (77 vs 72). But Inkling Small comes close despite being smaller, which is the notable part of the comparison.
What does “active parameters” mean and why does it matter?
In mixture-of-experts models like Longcat 2.5, only a subset of the total parameters are used for any given inference pass. Longcat has 1.6 trillion total parameters but only 48 billion active, meaning inference speed and compute cost track closer to the active number, while storage and memory requirements track the total.
Are these models free to use?
The weights themselves are open and downloadable, meaning you can self-host them without paying a usage fee. If you access them through a hosted API instead of running them yourself, providers charge per token, as with Ripple’s listed pricing of 20 cents per million input tokens and 60 cents per million output tokens.
