Inkling
Inkling is Thinking Machines Lab's first open-weight foundation model — a 975B-parameter sparse Mixture-of-Experts multimodal transformer (41B active parameters, released July 15, 2026) that reasons over text, images, and audio with a controllable "thinking effort" dial. Built for general-purpose reasoning, coding, and agentic/tool-use workloads, it delivers broad, balanced capability with strong open-weights standing (e.g., #1 US open-weights model on Artificial Analysis) at ~210 tokens/sec hosted throughput.
Empirical Evaluation Results
Architectural Profile & Capabilities
Inkling (OpenRouter slug thinkingmachines/inkling) is the debut open-weight release from Thinking Machines Lab, published July 15, 2026. It is a sparse Mixture-of-Experts (MoE) transformer with 975B total parameters and 41B active parameters per token, pretrained on ~45 trillion multimodal tokens spanning text, images, audio, and video. Unlike many competitive frontier variants, Inkling accepts text, image, and audio inputs and emits text, and exposes a variable/controllable "thinking effort" mechanism so operators can trade reasoning-token spend against latency and quality. It is positioned as a broad, balanced generalist rather than a benchmark-specialist, excelling more on agentic and multimodal tasks than on raw factuality. Positioning vs. the October 2026 frontier: Inkling does not try to match closed flagships (GPT-6 Astra, Claude Opus 5, Gemini 3.8 Pro) on absolute reasoning ceilings — its AA Intelligence Index of 25.0 and AA-Omniscience Index of ~2 are modest — but it is a leading *open-weights* option, ranked as the top U.S. open-weights model on Artificial Analysis when released and comparing favorably with open leaders such as Kimi K2.x, DeepSeek V4 Pro, and GLM on agentic and design-oriented evaluations. Its value proposition is open weights + multimodal reasoning + efficient active-parameter compute + controllable thinking, at aggressive pricing ($1/1M input, $4.05/1M output). Context window is deployment-dependent: Thinking Machines' general product page cites up to 1M tokens, the Tinker platform lists a 256K maximum, and the OpenRouter/provider metadata (and Benchmark Heaven) expose 524,288 tokens. Treat 524,288 as the operative served limit unless a deployment is documented otherwise. Output speed measured on Benchmark Heaven is ~210 tokens/sec, which for a 975B-class MoE is high-speed hosted performance. Benchmark integrity note: all scores below are ground-truth entries from Benchmark Heaven / admin calibration (carried forward unaltered) plus a small set of additional, explicitly Inkling-attributed figures surfaced from the Artificial Analysis release article and the OpenRouter Inkling benchmark panel (GDPval-AA v2 Elo, AA-Omniscience accuracy/non-hallucination, and Design Arena category Elos). No sibling, family, predecessor, or proxy scores (e.g., no Inkling-Small figures) have been substituted, and no scores were estimated.
975B total parameters with 41B active parameters per token; transformer decoder with MoE routing over feed-forward experts (expert count, experts-per-token, and routing scheme not publicly disclosed). Multimodal encoder fusion for text, image, and audio inputs with text-only outputs. Supports a controllable "thinking effort" dial governing test-time reasoning token budget. Pretrained on ~45T multimodal tokens (text, images, audio, video). Context window is provider/deployment-dependent: official product page cites up to 1M tokens, Tinker lists 256K, and OpenRouter/Benchmark Heaven expose 524,288; attention mechanism variant (e.g., GQA/MLA) is not officially specified. Open weights; API access via Tinker and third-party inference providers.
Recommended Workloads & Primary Use Cases
- •Deployment-dependent context: official page cites 1M, Tinker lists 256K, and OpenRouter/Benchmark Heaven expose 524,288 tokens — verify the served limit and long-context fidelity per provider; MLCR (12.2%) and GDP PDF (12.8%) indicate real long-context/document degradation.
- •Highly benchmark/suite-sensitive agentic behavior: Terminal-Bench 2.1 scores 55.1% while Terminal-Bench 4.0 collapses to 1.0%, and AutomationBench is only 5.0% — harness version, tool schema, and prompt scaffolding materially change results, so validate on your own agent stack.
- •Weak factual recall / elevated hallucination risk: AA-Omniscience Index ~2 and AA Intelligence Index 25.0 signal that grounding, RAG, and citation-verification layers are required for knowledge-intensive tasks.
- •Controllable thinking effort means latency and cost are non-deterministic: reasoning-token budgets vary by task, so TTFT and output pricing can swing; pin an effort setting and enforce output-token ceilings in production.
- •Open weights at 975B total parameters require substantial serving infrastructure; hosted pricing ($1/1M input, $4.05/1M output) plus variable reasoning tokens can raise effective cost on reasoning-heavy prompts.
Calls to Inkling are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Thinking Machines.
More Models from Thinking Machines
Compare other engines available in this laboratory.