What are the strengths of this model?

Strong logical reasoning and problem-solving capabilities. Designed for structured thinking and analysis tasks. Effective at breaking down complex problems into steps.

What are the weaknesses of this model?

May lack extensive general knowledge outside reasoning domains. Performance on creative or open-ended tasks might be limited. Dependent on clear input for optimal reasoning chains.

What are the best use cases?

Complex problem solving and logical deduction tasks. Mathematical proofs and technical analysis. Workflow planning and strategic decision support.

Thinking Machines LabOpen Source

Inkling

Name: Inkling
Author: Thinking Machines Lab

Compare this model

Inkling is Thinking Machines Lab's first train-from-scratch model, released July 15, 2026 under the Apache 2.0 license. A 975B-parameter sparse Mixture-of-Experts (41B active per token) with a 1,000,000-token context window, it is the leading U.S. open-weights model on the Artificial Analysis Intelligence Index (score 41). It accepts text, image, and audio input and generates text, trained on ~45 trillion tokens across text, image, audio, and video. Positioned as a customizable fine-tuning base rather than a finished chat product, it scores 77.6% on SWE-Bench Verified and 97.1% on AIME 2026. Weights are free to self-host; the Tinker API runs $1.87/$4.68 (64K) to $3.74/$9.36 (256K) per million tokens.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

License

Apache 2.0

Release Date

2026-07-15

Japanese Language Capability

🌐Multilingual

General multilingual model. Basic Japanese processing is possible, but inferior to specialized models.

API Pricing

Input Price (per 1M tokens)

$3.74

Output Price (per 1M tokens)

$9.36

Billing Mode: standard

Strengths

•Strong logical reasoning and problem-solving capabilities.
•Designed for structured thinking and analysis tasks.
•Effective at breaking down complex problems into steps.

Weaknesses

•May lack extensive general knowledge outside reasoning domains.
•Performance on creative or open-ended tasks might be limited.
•Dependent on clear input for optimal reasoning chains.

Use Cases

•Complex problem solving and logical deduction tasks.
•Mathematical proofs and technical analysis.
•Workflow planning and strategic decision support.

Deep Analysis

Total/Active Parameters

975B/41B

Mixture-of-Experts architecture

Context Window

1M tokens

Available via open weights

SWE-Bench Verified

77.6%

vs. Nemotron 3 Ultra: 70.7%

GPQA Diamond

87.2%

Strong science reasoning

VoiceBench

91.4%

Top-tier open-weights audio performance

Input Price (64K ctx)

$1.87/1M

Via Tinker API

FORTRESS Adversarial

78.0%

Best open-weights safety

Strengths

・Native multimodal reasoning over text, image, and audio with a unified architecture
・Controllable thinking effort (0.2-0.99) for precise cost-performance optimization
・Strong agentic and coding capabilities with excellent token efficiency
・Apache 2.0 license enabling unrestricted enterprise customization and deployment
・Robust safety profile with best-in-class refusal of harmful queries among open models

Weaknesses

・Factuality scores (SimpleQA: 43.9%) significantly trail leading closed models like GPT-5.6 Sol (71.6%)
・Massive hardware requirements (2TB VRAM for BF16) limit accessible local deployment
・Peak reasoning (HLE text-only: 29.7%) is outperformed by specialized models like GLM 5.2 (40.1%)
・Creative writing and nuanced instruction following lag behind frontier closed models

Competitor Comparison

Model	Arena	SWE	GPQA	Price
GLM 5.2	N/A	80.0%	89.5%	Open Weights
DeepSeek V4 Pro	N/A	80.6%	88.8%	Open Weights
Claude Fable 5 (max)	N/A	95.0%	92.6%	Closed Weights
GPT-5.6 Sol (xhigh)	N/A	82.2%	94.1%	Closed Weights

Overview

Inkling is Thinking Machines Lab's inaugural open-weights foundation model, designed as a broad, balanced generalist optimized for customization rather than benchmark dominance. With 975B total parameters (41B active) via a sparse Mixture-of-Experts architecture, it natively processes text, images, and audio through a unified encoder-free design. The model's key innovation is its controllable thinking effort, allowing developers to dial the reasoning budget from 0.2 to 0.99 to precisely balance performance against token cost and latency. Positioned as a practical multimodal foundation for enterprise fine-tuning, Inkling achieves competitive but not state-of-the-art performance across reasoning, coding, and safety benchmarks while offering unique advantages in native multimodality, token efficiency, and true open-source licensing (Apache 2.0). It excels in agentic workflows and represents the leading open-weights model from a U.S. lab, though it trails specialized Chinese open models on pure reasoning and top closed models on peak capabilities. The release signals Thinking Machines' focus on building customizable, efficient, and trustworthy AI systems rather than pursuing frontier scale alone. The model is available for immediate fine-tuning via the company's Tinker platform, with full weights on HuggingFace and partnerships with major inference providers for deployment.

Benchmarks & Performance

Inkling delivers a strong, balanced performance profile across diverse benchmarks at effort=0.99: ### Reasoning & Knowledge - **HLE (text-only):** 29.7% (vs. GLM 5.2: 40.1%, Claude Fable 5: 53.3%) - **HLE (with tools):** 46.0% (vs. Kimi K2.6: 54.0%, Gemini 3.1 Pro: 51.4%) - **AIME 2026:** 97.1% (vs. GPT-5.6 Sol: 99.9%, DeepSeek V4 Pro: 96.7%) - **GPQA Diamond:** 87.2% (vs. GPT-5.6 Sol: 94.1%, Kimi K2.6: 91.1%) ### Agentic Coding - **SWE-Bench Verified:** 77.6% (vs. Nemotron 3 Ultra: 70.7%, DeepSeek V4 Pro: 80.6%) - **Terminal Bench 2.1:** 63.8% (vs. GLM 5.2: 82.7%, GPT-5.6 Sol: 89.5%) - **GDPval-AA v2 (Elo):** 1238 (vs. Kimi K2.6: 1190, Nemotron 3 Ultra: 1164) ### Multimodal - **MMMU Pro (Standard 10):** 73.5% (vs. Gemini 3.1 Pro: 82.0%) - **VoiceBench:** 91.4% (vs. Gemini 3.1 Pro: 94.3%) - **MMAU:** 77.2% (vs. Gemini 3.1 Pro: 82.5%) ### Factuality & Safety - **SimpleQA Verified:** 43.9% (vs. GPT-5.6 Sol: 71.6%, DeepSeek V4 Pro: 57.0%) - **FORTRESS (Adversarial):** 78.0% (vs. Nemotron 3 Ultra: 77.6%, DeepSeek V4 Pro: 36.0%) - **StrongREJECT:** 98.6% (vs. Kimi K2.5: 99.5%) The model achieves a 77.1% on BrowseComp with context management, demonstrating strong factual grounding when augmented with search. Its controllable thinking effort is a key differentiator, achieving equivalent performance to Nemotron 3 Ultra on Terminal Bench 2.1 with roughly one-third of the tokens.

Detailed Comparison

### vs. GLM 5.2 (Open-Weights Leader) - **Performance:** GLM 5.2 leads on pure reasoning (HLE text: 40.1% vs. 29.7%) and agentic coding (Terminal Bench: 82.7% vs. 63.8%). Inkling offers native multimodality and better token efficiency. - **Pricing:** Both are open-weights. Inkling available via Tinker at $1.87/$4.68 per 1M tokens (input/output). - **Use Case:** Choose GLM for maximum reasoning/coding performance; choose Inkling for multimodal applications and cost-optimized deployment. ### vs. DeepSeek V4 Pro (Open-Weights Rival) - **Performance:** DeepSeek has higher factuality (SimpleQA: 57.0% vs. 43.9%) and coding (SWE-Bench: 80.6% vs. 77.6%). Inkling leads in mathematical reasoning (AIME: 97.1% vs. 96.7%) and safety (FORTRESS adversarial: 78.0% vs. 36.0%). - **Architecture:** Inkling is natively multimodal; DeepSeek V4 Pro appears text-focused in provided benchmarks. - **Differentiator:** Inkling's controllable thinking effort provides fine-grained cost control unavailable in DeepSeek's offering. ### vs. Claude Fable 5 (Closed-Weights Frontier) - **Performance:** Claude dominates on peak tasks (SWE-Bench: 95.0%, HLE text: 53.3%) but is closed-source and more expensive. - **Pricing:** Inkling is ~95% cheaper at scale ($1.87 input vs. closed-source enterprise pricing). - **Capabilities:** Inkling offers native audio processing absent from most closed models. Claude has superior creative nuance and complex instruction following. - **Positioning:** Inkling serves as a customizable, cost-efficient alternative for enterprises needing controlled deployment.

Community Feedback

Initial reception highlights both technical admiration and strategic positioning: - **Engineering Acclaim:** Researchers praise the full-stack training effort. Horace He (Thinking Machines) noted: "It truly takes a village to release a model, perhaps especially an open weights model. Actually doing the entire process from scratch... gives a lot of appreciation for anyone who does it!" - **Ecosystem Integration:** Hugging Face's Lysandre Jekinson highlighted inference optimizations: "We replaced the model's causal Conv1D with the `causal-conv1d` kernel. One line changed, +4% tokens per second... That's a total throughput improvement of about 15%." - **Open-Source Celebration:** Developers welcome the Apache 2.0 licensing. Tiezhen Wang emphasized its "975B total, 41B active" parameters and "Native MTP support" as wins for the open community. - **Use-Case Validation:** Early adopters on Tinker are exploring fine-tuning for domain-specific applications. The model's ability to "fine-tune itself" (as demonstrated in the release announcement) showcases unique self-improvement potential. - **Strategic Positioning:** Industry observers note the model positions Thinking Machines as a provider of "customizable foundations" rather than pursuing state-of-the-art at all costs. The focus on epistemics and censorship resistance resonates with users concerned about model biases and over-refusal. - **Benchmark Scrutiny:** Some community discussion compares Inkling to Chinese open models, noting that while U.S.-leading, it trails GLM 5.2 on pure reasoning—a trade-off acknowledged by the lab as intentional for breadth and multimodality.

Use Cases

1. **Agentic Coding Assistants:** Inkling excels in IDE-integrated coding agents where cost and latency matter. Its controllable thinking effort allows simple completions at low effort (minimal tokens) and complex debugging at high effort. The 77.6% on SWE-Bench Verified makes it suitable for generating production-quality code patches. Example: A developer fine-tunes Inkling on their proprietary codebase to create a context-aware assistant that understands internal APIs and style guides. 2. **Multimodal Customer Support Systems:** Native audio and image understanding enables building support bots that can process screenshots of error messages, transcribe customer voice calls, and analyze product images. Example: A retail company fine-tunes Inkling to handle returns by understanding photos of damaged items, spoken complaints, and order histories simultaneously. 3. **Cost-Sensitive Research & Education:** The model's strong reasoning at adjustable effort levels makes it ideal for educational tools where queries range from simple factual questions to complex problem-solving. Example: A university deploys Inkling as a teaching assistant that provides basic explanations at low effort and detailed step-by-step solutions at high effort, all within a fixed budget. 4. **Custom Enterprise Knowledge Agents:** With the Apache 2.0 license and fine-tuning on Tinker, organizations can build specialized models that incorporate proprietary data while maintaining full control. Example: A law firm fine-tunes Inkling on legal documents to create a research assistant that can reason over case law, contracts, and compliance materials without data leaving their environment. **When to choose Inkling over alternatives:** - Choose Inkling over closed models when data privacy, customization, or cost are primary concerns. - Choose over specialized open models (like GLM 5.2) when multimodality, token efficiency, or safety balance are critical. - Choose over smaller models when the task requires 1M token context or complex cross-modal reasoning. - Avoid when the application requires state-of-the-art creative writing or the highest possible accuracy on pure reasoning tasks.

Latest News

- **July 15, 2026:** Official release of Inkling (975B/41B active parameters) and Inkling-Small preview (276B/12B active) under Apache 2.0 license. Weights available on HuggingFace. - **July 15, 2026:** Day-0 support in Transformers (v5.14.0), SGLang, vLLM, and llama.cpp via Unsloth quantizations. - **July 15, 2026:** Launched on Tinker platform for fine-tuning with a limited-time 50% discount. Inkling Playground now available in Tinker console. - **July 15, 2026:** Partnerships for inference deployment announced with TogetherAI, Fireworks, Modal, Databricks, Baseten, and others. - **July 15, 2026:** NVFP4 checkpoint released optimized for NVIDIA Blackwell systems, reducing VRAM requirements from 2TB to 600GB. - **July 15, 2026:** Published detailed model card and technical report covering architecture, training, and safety evaluations. - **July 15, 2026:** Artificial Analysis ranks Inkling as #41 on Intelligence Index, making it the leading U.S. open-weights model. - **Ongoing:** Inkling-Small testing in progress with full weights release planned soon. The company indicates continued development of the model family.

Positioned as a practical multimodal foundation for enterprise fine-tuning, Inkling achieves competitive but not state-of-the-art performance across reasoning, coding, and safety benchmarks while offering unique advantages in native multimodality, token efficiency, and true open-source licensing (Apache 2.0). It excels in agentic workflows and represents the leading open-weights model from a U.S. lab, though it trails specialized Chinese open models on pure reasoning and top closed models on peak capabilities.

The release signals Thinking Machines' focus on building customizable, efficient, and trustworthy AI systems rather than pursuing frontier scale alone. The model is available for immediate fine-tuning via the company's Tinker platform, with full weights on HuggingFace and partnerships with major inference providers for deployment.

Sources

Analysis generated: 2026-07-17