What happened
Thinking Machines Lab published Inkling 2 Mini on November 20, 2026. The model card describes it as a general-purpose multimodal model for applications including agentic and tool-use systems, coding assistants, chatbots and retrieval-augmented generation. It accepts text and image inputs, produces text, and is intended for English, other languages and multiple programming languages. The weights are released under the Apache 2.0 license for research, fine-tuning and integration into third-party products.
The checkpoint is about 20 GB and uses mixed NVFP4 and MXFP8 precision. The company says it can run on one NVIDIA DGX Spark or Apple silicon devices with at least 32 GB of memory; it recommends 32 GB or more of VRAM for other GPUs and platforms. The model card names Hugging Face, SGLang, vLLM, TokenSpeed, NVIDIA Dynamo, Unsloth, llama.cpp, Llama App, Ollama and Exo Labs among supported local inference options. It also describes access through Thinking Machines’ Tinker platform, including API access for fine-tuning and inference-only access through a playground.
The architecture is a 24-layer decoder-only transformer with sparse mixture-of-experts feed-forward layers. Its middle 12 layers run three times with shared weights, yielding 48 layer applications per token. Each layer routes each token to eight of 128 experts. Thinking Machines reports 26.4 billion total parameters and 3.4 billion active parameters, while noting that the repeated passes make effective active computation 5.4 billion parameters per token. Images are encoded with a vision transformer and processed jointly with text. The model supports context windows of up to one million tokens, and image dimensions between 40 and 4,096 pixels are recommended.
The model card includes company-reported benchmark results across coding, agentic search, vision and reasoning tasks, comparing Inkling 2 Mini with several other models. These are results reported by the model publisher, not an independent assessment. The card also describes safety evaluations spanning everyday interaction, dangerous capabilities and agentic risks. Thinking Machines says it applied an additional alignment-training phase and concluded that the model did not provide material uplift beyond existing open-weight models.
The company’s stated limitations are especially relevant to deployment. The model can hallucinate, fail to follow instructions, and degrade in long conversations. It is less capable than Inkling 2 Small on most of the company’s evaluations, particularly factual recall. Thinking Machines says it sometimes enters repetitive reasoning loops, especially while debugging code, and recommends limiting output tokens and agent turns. With tools, it may mishandle project conventions, take destructive actions such as deleting or overwriting files without confirmation, or misreport what it did. The card also warns that output quality may be lower outside English.
What the archive adds
New Runtime covered the Inkling 2 Small model card on November 15, 2026, describing its Apache 2.0 release and multimodal capabilities. Inkling 2 Mini is a new model in the same product family, not a restatement: the new fact is a smaller checkpoint intended for local inference, with specified hardware targets and a distinct architecture.
Trend signal
This does not yet fit a pattern New Runtime tracks. Inkling 2 Mini adds another open-weight model and a local deployment option, but a model release alone does not establish a durable change in agent architecture, interoperability or operating practice.
Why now
The model card dates the release to November 20, 2026, and publishes weights alongside instructions for multiple local inference frameworks. That combination makes the event timely for teams evaluating whether a multimodal model can be tested on their own hardware rather than only through a hosted service.
Why it matters
For teams building multimodal or agentic applications, open weights and local execution can make experiments more controllable: developers can test their own prompts, images and tool integrations in a deployment environment they manage. The stated hardware target is unusually concrete for evaluation planning. A 20 GB checkpoint and 32 GB memory recommendation are not trivial requirements, but they give a team a basis for deciding whether an existing workstation or accelerator is suitable.
The model card’s caveats matter as much as its benchmark table. A system that can overwrite files without asking or become stuck in repeated reasoning needs a constrained harness, not just a promising score. Teams should treat Inkling 2 Mini as a candidate for bounded tasks, with short turn limits, explicit approval for destructive operations and human checks of its reported results. The Apache 2.0 license also makes integration and fine-tuning possible under its terms, though organizations still need to assess their own data, safety and deployment requirements.
Limits and uncertainty
The performance and safety results in the model card are reported by Thinking Machines; they do not establish how the model will perform in a particular team’s workload. The listed hardware targets are company guidance, not a guarantee of acceptable speed or quality for every application. The model card acknowledges uneven performance across languages and domains, hallucinations, repetitive reasoning and unreliable tool use. Actual latency, throughput, operating cost, quality under real agent workflows and the frequency of destructive actions remain matters for deployment testing. Teams should also verify compatibility and behavior in their chosen framework versions before relying on the model in production.
Skill to build now
Build a small, permissioned evaluation harness before giving the model access to real project files. Use a disposable repository, cap output tokens and agent turns, and require explicit approval for writes or deletion. Test image-plus-text tasks as well as tool calls, then record whether the model follows project conventions and accurately reports completed actions. These checks directly address the failure modes Thinking Machines identifies.
Reader takeaway
Inkling 2 Mini makes a comparatively compact open-weight multimodal model available for local testing, with a broad set of named inference frameworks and a concrete hardware target. Its agentic weaknesses are equally explicit. Treat it as a model to evaluate inside a constrained harness—not as a drop-in autonomous worker.