---
type: "post"
id: "inkling-2-small-model-card-826206"
slug: "inkling-2-small-model-card-826206"
title: "Inkling 2 Small Model Card"
description: "Thinking Machines Lab released Inkling 2 Small, an Apache 2.0 open-weight multimodal model with 594 billion total parameters, 17 billion active per token, and a context window of up to one million tokens."
retrieval_nugget: "Thinking Machines Lab released Inkling 2 Small, an Apache 2.0 open-weight multimodal model with 594 billion total parameters, 17 billion active per token, and a context window of up to one million tokens. Its permissive license broadens experimentation, but self-hosting still requires a large GPU cluster, so the practical choice is between hosted access and substantial infrastructure."
published_at: "2026-11-14"
updated_at: "2026-11-14"
record_date: "2026-11-14"
date_kind: "published_at"
topics: ["multimodal"]
entities: ["Thinking Machines Lab","Inkling 2 Small","Tinker","Hugging Face","SGLang","vLLM","TokenSpeed","Unsloth"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
what_changed: null
prior_refs: []
source_url: "https://thinkingmachines.ai/model-card/inkling-2-small"
source_urls: ["https://thinkingmachines.ai/model-card/inkling-2-small"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:inkling-2-small-model-card-826206"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/inkling-2-small-model-card-826206/","markdown":"https://newruntime.com/posts/inkling-2-small-model-card-826206.md","json":"https://newruntime.com/posts/inkling-2-small-model-card-826206.json"}
markdown_url: "https://newruntime.com/posts/inkling-2-small-model-card-826206.md"
---

# Inkling 2 Small Model Card

## Retrieval answer

Thinking Machines Lab released Inkling 2 Small, an Apache 2.0 open-weight multimodal model with 594 billion total parameters, 17 billion active per token, and a context window of up to one million tokens. Its permissive license broadens experimentation, but self-hosting still requires a large GPU cluster, so the practical choice is between hosted access and substantial infrastructure.

## Summary
Thinking Machines Lab released Inkling 2 Small, an Apache 2.0 open-weight multimodal model with 594 billion total parameters, 17 billion active per token, and a context window of up to one million tokens. Its permissive license broadens experimentation, but self-hosting still requires a large GPU cluster, so the practical choice is between hosted access and substantial infrastructure.

## What happened
Thinking Machines Lab’s model card gives November 5, 2026 as the release date for Inkling 2 Small. The company describes it as a general-purpose multimodal autoregressive transformer intended for developers building AI-powered applications, including agentic and tool-use systems, coding assistants, chatbots, and retrieval-augmented generation. It accepts text and image input and produces text. The model card says it is intended for English and other languages and supports multiple coding languages.

The architecture is a 48-layer decoder-only transformer with a sparse mixture-of-experts feed-forward backbone. Each token is routed to eight of 384 experts. The model has 594 billion total parameters, with 17 billion active, and uses a hybrid of local and global attention layers. Images are encoded through a vision transformer; image and text inputs are projected into a shared hidden space and processed jointly. Its context window extends to one million tokens. Image dimensions between 40 and 4,096 pixels are described as optimal.

Thinking Machines distributes weights in NVFP4 and MXFP8 formats through Hugging Face. It recommends NVFP4, the precision used in pretraining, midtraining, and post-training. The company says MXFP8 shows a marginal 1–2 percentage-point gain over NVFP4 across evaluations, which it considers practically insignificant. Running the weights directly requires at least 500 GB of aggregate VRAM for NVFP4, with examples of two NVIDIA B300 GPUs or four H200 GPUs. MXFP8 requires at least 800 GB, exemplified by four B300s or eight H200s. The company names SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face among the deployment frameworks.

Other access routes are Tinker, which provides API access and fine-tuning, Tinker’s inference-only playground, and third-party inference providers. The model card states an Apache 2.0 license and says the open weights support research, fine-tuning, and integration into third-party products. Its evaluations cover coding, search and tools, knowledge work, visual tasks, STEM, safety, and other categories. The company reports, among other results, 93.0% on GPQA Diamond, 81.3% on Terminal Bench v2.1, 81.7% on ARC-AGI-2, and 90.2% on BrowseComp with compaction. It also describes safety testing across everyday interactions, dangerous capabilities, and agentic risks, and recommends that downstream teams add safeguards and evaluate performance for their own uses.

## What the archive adds
New Runtime’s earlier coverage on [Inkling 2 Small’s launch on Tinker](https://newruntime.com/posts/thinking-machines-launches-inkling-2-small-moe-on-396573/) on November 6, 2026 described the 594-billion-parameter model and the plan to release Apache 2.0 weights after external safety testing. This model card makes the weights and deployment details concrete: it documents the license, checkpoint formats, hosted access routes, and hardware requirements. This is a new fact on a known story, not a separate model launch.

## Trend signal
This release supports [agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/) only weakly: the model card reports results on agent-adjacent benchmarks, including coding and browsing, but does not establish cost per verified completed task. Its extensive benchmark table is a reminder that model rankings alone do not tell teams what a deployed agent will cost or how reliably it will finish work. The useful comparison remains task-level performance under a team’s own harness and constraints.

## Why now
The model card dates the release to November 5, 2026, while New Runtime’s earlier coverage the following day described the weights as planned after external safety testing. The current card identifies Apache 2.0 weights and provides concrete hardware and access information, making the release more actionable for teams deciding whether to test hosted access or self-hosting.

## Why it matters
For builders, the headline is not simply a high parameter count or a long context window. The release combines an open license with multiple deployment routes, making it possible to experiment through a hosted API before committing to a cluster or to fine-tune and integrate weights into a product. That flexibility can matter for teams that need control over model behavior or deployment.

The local hardware threshold changes the economics: 500 GB of aggregate VRAM for the recommended NVFP4 checkpoint is a serious operational requirement, and the alternative MXFP8 format raises it further. The model card’s benchmark results can guide which tasks to test, but they are reported by the model’s developer and do not establish production quality, latency, or cost for a particular workload. Teams should compare the hosted and self-managed routes against the same prompts, tools, and verification criteria before choosing.

## Limits and uncertainty
The model card contains extensive evaluations and safety findings reported by Thinking Machines, but those claims are not independently verified here. Benchmark scores do not establish performance in a team’s own agent harness, language mix, or production conditions. The card itself identifies possible hallucinations, instruction-following failures, degraded long-conversation performance, uneven language or domain performance, and training-cutoff limits. It recommends human review for high-stakes uses and additional application-layer safeguards. Actual inference cost, throughput, and operational reliability will depend on deployment and are not determined by the reported benchmark scores.

## Reader takeaway
Inkling 2 Small is a genuinely open-weight release with useful hosted and self-hosted routes, but “open” does not mean lightweight: local deployment starts at a large multi-GPU footprint. Treat the company’s evaluations as a shortlist for your own tests, and measure completed-task quality, cost, and safety on the workloads you intend to run.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)