The model features 26.4 billion total parameters with 3.4 billion active parameters, cycling middle layers to produce 5.4 billion weight activations. While this looping design boosts serial problem-solving, it does not expand underlying factual storage, prompting the system to rely heavily on search tools to match larger competitors. Native deployment utilizes NVFP4 quantization, while its final four layers operate in MXFP8. Developers can also evaluate a variable-loop experimental version to explore deeper multi-pass computation.
Key points
Looped transformers can drastically boost serial reasoning and agentic coding performance…: Looped transformers can drastically boost serial reasoning and agentic coding performance without expanding active memory requirements.
Because layer looping does not improve internal factual storage, the system compensates b…: Because layer looping does not improve internal factual storage, the system compensates by leveraging external search and tool integration.
Hardware efficiency is achieved through NVFP4 deployment with the final four layers runni…: Hardware efficiency is achieved through NVFP4 deployment with the final four layers running in MXFP8, enabling offline execution on consumer laptops.
Sources checked by our analysis pipeline: 1
Today we are releasing Inkling 2 Mini, a fast, powerful model that’s small enough to run on your laptop.
Inkling 2 Mini is a looped transformer, running the middle layers of the model three times for every token. This allows the model to apply more reasoning and reach the right answer with fewer tokens, especially on tasks such as agentic coding that depend on reasoning quality. Inkling 2 Mini leads every open-weights model with less than 100B total parameters on benchmarks such as DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). It runs faster than similarly performing models as well.
Inkling 2 Mini is a Mixture-of-Experts model with 26.4B total parameters, 3.4B active, and 5.4B weight activations per token to account for looping. Looping allows the model to reason longer without writing its reasoning down, which matters for monitoring: our ability to oversee and understand what the model does by reading its chain of thought (CoT).
We measured the impact of looping on both performance and monitorability with Inkling 2 Mini and a research-specific variable-loop variant we are releasing alongside…
Software teams can now deploy an autonomous coding agent directly inside their own infrastructure or through a managed cloud platform. Operated via simple Slack mentions, the tool executes development tasks on an isolated machine and delivers results as pull requests. This release provides organizations with an open, self-hosted option to accelerate automated software engineering without locking their workflows into proprietary services.
Bad actors are increasingly attempting to extract advanced frontier model capabilities to create unaligned, guardrail-free systems capable of biological or cyber harm. OpenAI has deployed comprehensive countermeasures across its infrastructure to prevent this unauthorized transfer, emphasizing the protection of underlying reasoning traces. These defenses aim to block malicious developers from weaponizing defensive research and dual-use technological discoveries.
Artificial intelligence company Anthropic appears to have solved the Riemann Hypothesis, resolving what has stood as mathematics' crown jewel since Fermat's Last Theorem was settled in the 1990s. Because countless conjectures were previously proven conditional on this single hypothesis, the machine-generated breakthrough instantly unlocks an immense body of mathematical research. However, the historic breakthrough has sparked bittersweet mourning rather than celebration, as observers view it as a signal that human…