Inkling

Overview

Inkling is the first model trained from scratch by Thinking Machines Lab, the research company founded by former OpenAI CTO Mira Murati, and its first open-weights release. Announced on July 15, 2026, it arrives as a broad, balanced generalist rather than a benchmark specialist — a single model meant to reason over text, images and audio, write and run code, drive a browser, and hold calibrated confidence about what it does and does not know.

Architecturally it is a large Mixture-of-Experts transformer: 975B total parameters with 41B active per token, a context window of up to 1M tokens, and pretraining on 45 trillion tokens of text, images, audio and video. The multimodality is native — audio and vision are reasoned over directly rather than bolted on through separate encoders. Alongside the flagship, Thinking Machines shipped a preview of Inkling-Small (276B total, 12B active), trained with a similar recipe for lower cost and latency, which matches or beats the larger model on several evaluations.

What makes the release notable is less any single score than its positioning. Thinking Machines published the full weights under Apache 2.0, making Inkling freely downloadable, modifiable and commercially usable, and wired it into its Tinker fine-tuning platform on day one. Artificial Analysis called it the leading U.S. open-weights model at release. For teams that have been reaching for Chinese open-weights models like DeepSeek V4 or Kimi K3 because the strongest open weights came from there, Inkling is the first competitive Western alternative in that tier.

Capabilities

• Native multimodal reasoning: Processes and reasons over text, images and audio in a single model, without separate modality adapters. It posts strong vision (MMMU Pro 73.5%) and audio (VoiceBench 91.4%) scores rather than treating those modes as afterthoughts.
• Controllable thinking effort: A first-class control lets you dial reasoning depth up or down to trade accuracy against latency and token spend. Thinking Machines reports frontier agentic results at ~25K output tokens per task, well below the 37–43K competitors spend.
• Agentic coding and tool use: Handles multi-step software tasks, browser automation and web-app generation, scoring 77.6% on SWE-bench Verified and 63.8% on Terminal-Bench 2.1.
• Advanced reasoning and math: 97.1% on AIME 2026 and 87.2% on GPQA Diamond place it near frontier closed models on hard reasoning and graduate-level science.
• Calibrated, epistemically honest answers: Thinking Machines emphasizes factual grounding and calibrated confidence — the model is tuned to express appropriate uncertainty rather than confidently hallucinate.
• Long-context work: Up to 1M tokens supports whole-repository reasoning, long documents and extended agent trajectories.

Technical Specifications

• Architecture: Mixture-of-Experts transformer
• Parameters (Inkling): 975B total, 41B active per token
• Parameters (Inkling-Small): 276B total, 12B active per token (preview)
• Modalities: Text, image and audio input; text output (native multimodal)
• Context window: Up to 1M tokens (open weights); 64K and 256K options on the Tinker API
• Pretraining data: 45 trillion tokens of text, images, audio and video
• Thinking: Controllable thinking effort; benchmark table measured at effort 0.99
• License: Apache 2.0
• Checkpoints: Original weights and an NVFP4 quantized checkpoint on Hugging Face
• Knowledge cutoff: Not disclosed
• Measured throughput: ~72.7 output tokens/sec with ~1.75s time-to-first-token on Thinking Machines' API (Artificial Analysis)

Use Cases

• Cost-controlled agents: The thinking-effort dial and low token spend per task make Inkling attractive for agent loops where every step's tokens are billed and latency compounds.
• Self-hosted deployment: Apache 2.0 weights let regulated or privacy-sensitive teams run the model on their own infrastructure with no per-token API fees and no data leaving their environment.
• Domain fine-tuning: First-party support on Tinker makes Inkling a base for fine-tuning on proprietary data, with 64K and 256K context recipes.
• Multimodal pipelines: Native audio and vision suit document understanding, chart and figure reasoning, and voice interfaces without stitching together separate models.
• Agentic software engineering: Repository-scale coding, terminal-driven tasks and browser automation, where its SWE-bench and Terminal-Bench results are competitive with far more expensive closed models.
• Research and evaluation baselines: An open, competitive Western model gives labs a reproducible baseline they can inspect, probe and modify.

Performance / Benchmarks

The figures below are vendor-reported, transcribed from Thinking Machines' launch announcement, measured at thinking effort 0.99. As with any first-party benchmark table, the developer chose the evals, harnesses and settings — treat these as an upper bound and validate on your own tasks.

A recurring pattern worth noting: Inkling-Small keeps up with, and sometimes beats, the flagship — it edges ahead on GPQA Diamond, IFBench and HLE-with-tools — while falling behind mainly on the hardest agentic tasks like Terminal-Bench 2.1 (52.7% vs 63.8%). For many workloads the smaller model may be the better default.

Artificial Analysis

On the Artificial Analysis Intelligence Index, Inkling scores 41 (v4.1) — reported as the leading U.S. open-weights model at release, three points above Nvidia's Nemotron 3 Ultra (38) and well ahead of Gemma 4 31B (29) and gpt-oss-120b (24). Its closest open competitors on that index remain the frontier Chinese models. On the GDPval-AA v2 agentic benchmark, Artificial Analysis measured an Elo of 1238 — above Kimi K2.6 (1190) and DeepSeek v4 Flash (1189) — while using only ~25K output tokens per task versus 37–43K for competing models, underscoring its efficiency claim.

How to read these numbers

Scores are not comparable across Intelligence Index versions (v4.0 and v4.1 can differ by more than ten points on the same model), so the v4.1 label matters. The launch table's effort setting (0.99) is near the top of Inkling's range; results at lower effort will be lower and cheaper. And first-party benchmark tables reflect the developer's chosen configuration — the honest way to size Inkling for your use is your own evaluation set.

Limitations

• Above-average price for an open-weights model: At roughly $1.87 input / $4.68 output per million tokens on the flagship, Inkling is priced above several open competitors; Artificial Analysis flags it as carrying above-average pricing for the tier despite the open license.
• Serving cost of a 975B model: Even at 41B active parameters, self-hosting a 975B-parameter MoE demands substantial GPU memory. The NVFP4 checkpoint eases this, but this is not a model that runs on a single consumer GPU.
• Text-only output: Inkling reasons over image and audio inputs but generates text, not images or audio.
• Undisclosed knowledge cutoff: Thinking Machines did not publish a training-data cutoff date, so time-sensitive factual coverage is hard to bound without testing.
• Vendor-reported benchmarks: The launch scores are first-party and measured at high effort; independent replication was still emerging at release.
• New model, thin ecosystem: As a day-one release, tooling, quantizations, community fine-tunes and deployment guides are less mature than for long-established open models like Llama 4 or DeepSeek.

Pricing & Access

Open weights (self-hosted) — Free to download and run under Apache 2.0. Full weights and an NVFP4 quantized checkpoint are on Hugging Face; commercial use, modification and redistribution are permitted. Your only cost is compute.

Thinking Machines Tinker API — Fine-tuning and inference are priced by context option:

At launch, Tinker fine-tuning of Inkling carried a 50% discount for a limited time.

Third-party hosted APIs — Inkling is available through TogetherAI, Fireworks, Modal, Databricks and Baseten, each setting its own price and rate limits.

Ecosystem & Tools

• Hugging Face: Original and NVFP4 checkpoints for self-hosting and further fine-tuning.
• Tinker: Thinking Machines' first-party platform for fine-tuning and serving Inkling, with 64K and 256K context recipes.
• Inference providers: TogetherAI, Fireworks, Modal, Databricks and Baseten offer hosted endpoints.
• Inkling-Small (preview): A lighter 12B-active variant for latency- and cost-sensitive deployments, sharing the training recipe.

Community & Resources

• Introducing Inkling — Thinking Machines Lab — official announcement
• Inkling on Artificial Analysis — independent index scores and pricing
• Artificial Analysis: the new leading U.S. open weights model — third-party analysis
• Tinker fine-tuning platform
• Apache 2.0 License