Thinking Machines Lab Inkling: 975B Open-Weights Multimodal MoE with Self-Improvement, Controllable Effort, and Apache 2.0 Freedom

Executive Summary

On July 15, 2026, Thinking Machines Lab — the startup founded by ex-OpenAI CTO Mira Murati alongside OpenAI co-founder John Schulman, former OpenAI VP Lilian Weng, and other ex-OpenAI leaders — released Inkling, its first open-weights foundation model. At 975 billion total parameters with 41 billion active per token, Inkling is a Mixture-of-Experts (MoE) transformer that natively processes text, images, and audio with a 1M-token context window. The model is licensed under Apache 2.0, making it one of the most permissively licensed frontier-class open models available.

What distinguishes Inkling from other recent open-weight releases is not raw benchmark dominance — the company explicitly states it is "not the strongest overall model available today, open or closed" — but a combination of qualities: native multimodality (especially audio, where it leads open-weight models), controllable thinking effort that lets developers trade latency and cost against performance, and a self-improvement capability demonstrated by having the model fine-tune itself via the Tinker platform. The company also previewed Inkling-Small, a 276B-parameter variant with only 12B active parameters that matches or exceeds the flagship on several benchmarks.

The release positions Thinking Machines as a US-based counterweight to China's dominant open-weight ecosystem (Kimi K3, DeepSeek V4-Pro, GLM-5.2) while pursuing a different strategy: rather than chasing the largest parameter count, they're optimizing for customizability — a model designed to be fine-tuned for specific domains rather than deployed as a general-purpose API.

1. The Thinking Machines Story: From OpenAI Exiles to $12B Valuation

1.1 Founding Team

Thinking Machines Lab was founded in February 2025 by a team of former OpenAI leaders:

• Mira Murati — CEO, former OpenAI CTO (May 2022 – September 2024)
• John Schulman — Co-founder, OpenAI co-founder and former research lead
• Lilian Weng — Co-founder, former OpenAI VP of Research
• Barret Zoph — Co-founder, former OpenAI VP
• Luke Metz — Co-founder, former OpenAI researcher
• Soumith Chintala — CTO, former PyTorch creator

This is one of the most high-profile departures from OpenAI, occurring during the governance crisis of late 2024. The team chose to build independently rather than join competing labs.

1.2 Funding

The company raised $2 billion in seed funding in 2025 at a $12 billion valuation — the largest seed round in history at the time. Key investors include:

• Andreessen Horowitz (a16z)
• NVIDIA
• AMD
• Cisco
• Jane Street

Notably, the company had not shipped a product when this round closed. Investors were betting on the team and the thesis, not on traction.

1.3 The Thesis: Customization Over Generalization

Thinking Machines' stated mission is to "build AI that extends human will and judgment." Their strategic bet is that customizable open-weight models will outperform one-size-fits-all closed models for enterprise and specialized use cases. This contrasts with the approach of OpenAI, Anthropic, and Google, which optimize for the strongest possible general-purpose model and offer limited fine-tuning.

2. Architecture: 975B MoE with Native Multimodality

2.1 Core Specifications

Property	Value
Total parameters	975B
Active parameters	41B per token
Architecture	66-layer decoder-only transformer
MoE configuration	256 experts + 2 shared experts; 6 experts routed per token
Attention	Hybrid local and global layers
Context window	1M tokens (64K/256K on Tinker API)
Input modalities	Text, Image, Audio
Output modality	Text
Training data	45 trillion tokens (text, images, audio, video)
Numerics	BF16, MXFP8, NVFP4
License	Apache 2.0

2.2 MoE Design

Inkling uses a sparse MoE architecture with 256 specialists plus 2 shared experts active on every token. Each token is routed to 6 of the 256 experts, meaning approximately 2.3% of the total parameters are active per forward pass. This yields 41B active parameters from a 975B total — a sparsity ratio of roughly 23.8x.

The shared experts ensure that certain capabilities (e.g., basic language modeling, instruction following) are always available regardless of routing decisions, while the specialists allow the model to develop domain-specific expertise.

2.3 Native Multimodality

Inkling is natively multimodal, not a text model with added vision/audio adapters. All modalities are projected into a shared hidden space and processed jointly by the decoder:

• Images: Encoded via a four-layer hierarchical hMLP (hierarchical Multi-Layer Perceptron) patch encoder operating on 40×40 pixel patches. Optimal input range is 40px to 4096px per dimension.
• Audio: Encoded as dMel spectrograms (discrete Mel-frequency spectrograms) from 16kHz WAV input. Optimal audio length is within 20 minutes.
• Text: UTF-8 encoded, processed as standard token sequences.

This encoder-free architecture for audio and vision inputs is consistent with the company's "interaction models" design, which aims for real-time voice and vision collaboration.

2.4 Inkling-Small

Alongside the flagship, Thinking Machines previewed Inkling-Small:

Property	Inkling	Inkling-Small
Total parameters	975B	276B
Active parameters	41B	12B
Architecture	Same MoE recipe	Same MoE recipe
IFBench	79.8%	83.4%
Target use case	General purpose	Cost/latency-sensitive workloads

Inkling-Small's ability to match or beat the flagship on benchmarks like IFBench (83.4% vs 79.8%) suggests that the training recipe and post-training stack are the primary differentiators, not just parameter count. This is particularly relevant for workloads where cost and latency matter: coding, grading, and synthetic data generation.

3. Controllable Thinking Effort: A Novel Feature

3.1 What It Is

Inkling supports controllable thinking effort, a parameter that lets developers adjust the model's reasoning depth from 0.2 (minimal) to 0.99 (maximum). This is not simply a temperature or top-p setting — it directly controls how much "thinking" the model does before producing a response.

3.2 The Efficiency Curve

The key finding from Thinking Machines' evaluation is that Inkling achieves a given performance level at far fewer tokens than competing models:

• On Terminal Bench 2.1, Inkling matches Nemotron 3 Ultra's performance at roughly one-third of the tokens.
• The effort sweep shows a smooth performance curve across Terminal Bench 2.1, Humanity's Last Exam (HLE), and IFBench.

3.3 Practical Implications

This feature enables cost-tiered agentic pipelines:

1. Low effort (0.2-0.4): Handle routing, triage, simple QA, and data extraction
2. Medium effort (0.5-0.7): Coding tasks, document analysis, tool use
3. High effort (0.8-0.99): Complex reasoning, bug repair, multi-step planning

This is a single deployment with multiple budget tiers, eliminating the need to maintain separate models for different complexity levels.

4. Benchmarks: Where Inkling Stands

4.1 Reasoning & Knowledge

Benchmark	Inkling (0.99)	Nemotron 3 Ultra	Kimi K2.5	Kimi K2.6	GLM 5.2	DeepSeek V4-Pro	GPT-5.6 Sol	Claude Fable 5
HLE (text)	29.7%	26.6%	29.4%	35.9%	40.1%	35.9%	47.2%	53.3%
HLE (tools)	46.0%	37.4%	50.2%	54.0%	54.7%	48.2%	55.0%	64.5%
AIME 2026	97.1%	94.2%	95.8%	96.4%	99.2%	96.7%	99.9%	99.9%
GPQA Diamond	87.2%	86.7%	87.9%	91.1%	89.5%	88.8%	94.1%	92.6%

On reasoning benchmarks, Inkling sits in the mid-tier of open-weight models — above Nemotron 3 Ultra and Kimi K2.5 but below Kimi K2.6, GLM 5.2, and DeepSeek V4-Pro. Against closed-weight models (GPT-5.6 Sol, Claude Fable 5), there is a clear gap, which the company acknowledges.

4.2 Agentic Coding

Benchmark	Inkling (0.99)	Nemotron 3 Ultra	Kimi K2.5	Kimi K2.6	GLM 5.2	DeepSeek V4-Pro	GPT-5.6 Sol	Claude Fable 5
SWE-Bench Verified	77.6%	70.7%	76.8%	80.2%	80.0%	80.6%	82.2%	95.0%
SWE-Bench Pro	54.3%	46.4%	50.7%	58.6%	62.1%	55.4%	64.6%	80.0%
Terminal Bench 2.1	63.8%	56.4%	51.3%	71.3%	82.7%	64.0%	89.5%	84.6%
MCP Atlas	74.1%	44.7%	64.0%	68.1%	77.8%	73.2%	81.8%	83.3%

Inkling's coding performance is competitive among open-weight models, particularly on SWE-Bench Verified (77.6%) where it outperforms Nemotron 3 Ultra and Kimi K2.5. On Terminal Bench 2.1, it achieves 63.8% — solid but trailing the specialist coding models (GLM 5.2 at 82.7%, DeepSeek V4-Pro at 64.0%).

4.3 Multimodal: Audio & Vision

This is where Inkling truly distinguishes itself:

Benchmark	Inkling (0.99)	Qwen3-Omni	Nemotron-3Nano-Omni	Kimi K2.5	Kimi K2.6	Gemini 3.1 Pro
Audio MC	56.6%	24.3%	23.2%	—	—	66.8%
MMAU	77.2%	77.5%	76.7%	—	—	82.5%
VoiceBench	91.4%	88.8%	89.4%	—	—	94.3%
MMMU Pro	73.5%	60.0%	53.0%	75.0%	79.0%	82.0%
CharXiv RQ	78.1%	61.1%	63.6%	77.5%	80.4%	80.2%
CharXiv RQ (Python)	82.0%	—	—	78.7%	86.7%	89.9%

On audio benchmarks, Inkling is the strongest open-weight model by a significant margin — scoring 56.6% on Audio MC compared to the next-best open model (Qwen3-Omni) at 24.3%. On VoiceBench, it achieves 91.4%, approaching Gemini 3.1 Pro's 94.3%.

On vision benchmarks, Inkling is competitive but not leading among open-weight models, with Kimi K2.6 (79.0%) and Kimi K2.5 (75.0%) slightly ahead on MMMU Pro.

4.4 Epistemics: Forecasting & Calibration

Benchmark	Inkling	GPT-5.5	Claude Opus 4.8	Gemini 3.1 Pro	Grok 4.3	Kimi K2.6
ForecastBench (no search)	61.1	59.1	54.6	61.1	58.1	61.7
ForecastBench (with search)	63.7	64.7	58.6	64.3	63.2	—
Prophet Arena (Brier ↓)	0.1617	0.1598	0.1605	0.1594	0.1715	0.1675

Inkling's forecasting performance is notable: it matches or exceeds GPT-5.5 and Claude Opus 4.8 on ForecastBench without search, and is competitive with search. This reflects the company's explicit training for calibration using RL against proper scoring rules.

4.5 Safety

Benchmark	Inkling	Nemotron 3 Ultra	Kimi K2.5	Kimi K2.6	GLM 5.2	DeepSeek V4-Pro	Claude Fable 5	GPT-5.6 Sol
FORTRESS (Adversarial)	78.0%	77.6%	54.1%	65.6%	71.3%	36.0%	96.0%	82.4%
FORTRESS (Benign)	95.9%	90.5%	98.3%	97.2%	90.0%	98.5%	55.1%	98.1%
StrongREJECT	98.6%	98.7%	99.5%	99.8%	98.5%	98.6%	98.7%	98.5%

Inkling's safety performance is strong among open-weight models, particularly on FORTRESS Adversarial (78.0%) where it significantly outperforms DeepSeek V4-Pro (36.0%) and Kimi K2.5 (54.1%). However, it trails Claude Fable 5 (96.0%) and GPT-5.6 Sol (82.4%) on adversarial resistance.

5. Self-Improvement: The Closed Customization Loop

5.1 The Demo

The most distinctive demonstration at launch was Inkling fine-tuning itself. The process worked as follows:

1. A user asked Inkling (running inside the OpenCode harness) to create a model that never uses the letter "e" — a lipogram constraint that prompting alone cannot reliably achieve.
2. Inkling wrote its own fine-tuning job using the Tinker platform.
3. It generated synthetic training data for the lipogram task.
4. It ran the training process on Tinker.
5. It evaluated the result against the base model.
6. It switched to the improved weights.

The company reported that during this self-improvement process, Inkling's chain-of-thought reasoning became more concise over time, "dropping grammatical overhead while remaining comprehensible and leaving the final response unaffected."

5.2 Why This Matters

This is not just a product demo — it demonstrates a closed customization loop where the model can:

• Identify its own limitations
• Generate training data to address them
• Execute the training
• Evaluate the improvement
• Deploy the updated weights

This capability is unique among open-weight models and represents a step toward self-improving AI systems — though the company is careful to frame it as a customization tool rather than autonomous evolution.

6. Deployment & Implementation

6.1 Hardware Requirements

Checkpoint	VRAM Required	Configuration Options
BF16	≥ 2 TB aggregated	8× NVIDIA B300, or 16× NVIDIA H200
NVFP4 (W4A4)	≥ 600 GB	4× NVIDIA B300 (SM100+ required)
NVFP4 (W4A16)	≥ 600 GB	8× NVIDIA H200

The BF16 checkpoint requires significant infrastructure — 8 B300 GPUs or 16 H200s — placing it out of reach for most individual developers. The NVFP4 quantized checkpoint reduces requirements to 4 B300s or 8 H200s, which is more accessible for enterprise deployments.

6.2 Software Stack

Inkling is compatible with major inference frameworks:

• SGLang
• vLLM
• TokenSpeed
• Unsloth
• Hugging Face Transformers

6.3 Access Options

1. Tinker API: Fine-tuning and inference through Thinking Machines' own platform. Includes the Inkling Playground for developer testing.
2. Hugging Face: Full weights available at thinkingmachines/inkling under Apache 2.0.
3. Third-party inference providers: Available through partner platforms.

6.4 Implementation Example

# Using Tinker SDK for fine-tuning
from tinker import TinkerClient

client = TinkerClient(api_key="your-api-key")

# Fine-tune Inkling for a specific domain
job = client.fine_tune.create(
    model="inkling",
    training_file="domain_data.jsonl",
    hyperparameters={
        "epochs": 3,
        "learning_rate": 2e-5,
        "effort": 0.7,  # Set thinking effort for training
    },
    suffix="financial-analyst",
)

# Deploy the fine-tuned model
deployment = client.deployments.create(
    model=job.fine_tuned_model,
    effort_range=[0.2, 0.99],  # Allow dynamic effort adjustment
)

7. Comparison with Recent Open-Weight Releases

7.1 The Open-Weight Landscape (July 2026)

Model	Total Params	Active Params	License	Multimodal	Key Differentiator
Inkling	975B	41B	Apache 2.0	Text, Image, Audio	Self-improvement, controllable effort
Kimi K3	2.8T	~50B	Apache 2.0 (July 27)	Text, Image	Largest open model, frontier coding
DeepSeek V4-Pro	1.6T	~25B	MIT	Text	Strong coding, disaggregated inference
GLM-5.2	~1T	Unknown	Apache 2.0	Text, Image	Long-horizon agents, strong reasoning
MiniMax M3	428B	23B	Apache 2.0	Text	Self-evolving agent harness
Nemotron 3 Ultra	Unknown	Unknown	Apache 2.0	Text	NVIDIA-backed, strong reasoning

7.2 Strategic Positioning

Inkling occupies a unique niche:

• Not the largest (Kimi K3 at 2.8T dominates on scale)
• Not the strongest coder (DeepSeek V4-Pro and GLM-5.2 lead)
• But the most multimodal (especially audio)
• And the most customizable (self-improvement + Tinker platform)

This positioning aligns with Thinking Machines' thesis: the value is not in the base model but in the ability to adapt it to specific domains.

8. Integration with Prior Research

Inkling's release connects to several themes in our research:

• Open-weight frontier expansion: Follows the trajectory documented in Kimi K3 Open 3t Class Model Frontier Coding Agentic Knowledge Work 2026 07 20 and Deepseek V4 Pro Frontier Analysis 2026 04 24, where China has dominated open-weight scale. Inkling provides a US-based alternative.
• MoE architecture evolution: The 256-expert configuration with extreme sparsity continues the trend analyzed in Mixture Of Experts Sparse Models Explained and Dense Transformers Vs Sparse Moe Architecture 2026 04 20.
• Agentic coding economics: The controllable effort feature directly addresses the cost concerns raised in Agentic Coding Economics Roi Adoption 2026 05 18 by enabling cost-tiered deployments.
• Multimodal reasoning: The native audio capabilities extend the multimodal analysis in Gemini 3 5 Flash Frontier Agents Coding Flash Tier Cost 2026 07 17 and Qwen Robot Suite Embodied Ai Navigation Manipulation World Model 2026 06 19.
• Self-improvement: The closed customization loop touches on themes from Minimax M27 Self Evolving Agent Harness Open Weight Frontier 2026 07 16 regarding self-evolving systems.

9. Key Takeaways

1. Inkling is not trying to be the strongest model — it's trying to be the most customizable. This is a deliberate strategic choice that differentiates it from Kimi K3 and DeepSeek V4-Pro.

2. Audio is the killer feature — Inkling's audio benchmarks (56.6% Audio MC, 91.4% VoiceBench) significantly outperform other open-weight models, making it the best open option for voice-enabled applications.

3. Controllable thinking effort is a practical innovation — the ability to adjust reasoning depth from 0.2 to 0.99 enables cost-tiered deployments without maintaining multiple models.

4. Self-improvement is real but bounded — the closed customization loop is impressive but operates within a human-defined framework. It's a tool for faster iteration, not autonomous evolution.

5. Inkling-Small may be the smarter choice for many use cases — with 12B active parameters matching or beating the flagship on key benchmarks, it offers dramatically lower inference costs.

6. The Apache 2.0 license is a strategic advantage — unlike some open models with restrictive licenses, Inkling can be freely commercialized, modified, and redistributed.

7. The US open-weight ecosystem is diversifying — with Thinking Machines joining Moonshot (China), DeepSeek (China), and NVIDIA (Nemotron), the open-weight landscape is becoming more geographically diverse.

10. Future Directions

10.1 What to Watch

• Inkling-Small full release: The preview suggests strong performance at 12B active parameters. Full weights and benchmarks will clarify its production readiness.
• Tinker platform maturity: The value proposition depends on how easy it is to fine-tune Inkling for specific domains. Enterprise adoption will hinge on the Tinker developer experience.
• Interaction models: The company has previewed a system built for real-time voice and vision collaboration. Inkling is positioned as the reasoning backend for this system.
• Model family expansion: The company explicitly stated this is "the first in a family of models of different sizes." Expect more variants in the coming months.
• Self-improvement research: The closed customization loop could evolve into more sophisticated self-improvement capabilities, potentially competing with MiniMax's self-evolving agent harness.

10.2 Potential Impact

If Thinking Machines' thesis is correct — that customizable open-weight models will outperform closed general-purpose models for enterprise use — Inkling could become a foundation for domain-specific AI systems across industries. The combination of Apache 2.0 licensing, native multimodality, and the Tinker fine-tuning platform creates a stack that is difficult for closed-model providers to replicate.

However, the model's performance gap on reasoning and coding benchmarks relative to Kimi K3 and DeepSeek V4-Pro means it will need to prove that customization can close the gap in practice. The financial-judgment fine-tuning work the company references is a promising early signal, but broader validation is needed.

11. References & Resources

Official Sources

• Inkling Announcement — Thinking Machines Lab official blog
• Inkling Model Card — Full technical specifications and evaluations
• Inkling Landing Page — Product overview
• Tinker Cookbook — Fine-tuning guide
• Hugging Face: thinkingmachines/inkling — Model weights
• Training Data Documentation — Data provenance
• Model Acceptable Use Policy — Usage guidelines

Key Research Papers Referenced

• dMel: Speech Tokenization made Simple — Audio encoding method used by Inkling
• Three things everyone should know about Vision Transformers — hMLP architecture basis

Related Journal Articles

• Kimi K3 Open 3t Class Model Frontier Coding Agentic Knowledge Work 2026 07 20 — Kimi K3 analysis
• Minimax M27 Self Evolving Agent Harness Open Weight Frontier 2026 07 16 — MiniMax M2.7 self-evolving agent
• Deepseek V4 Pro Frontier Analysis 2026 04 24 — DeepSeek V4-Pro analysis
• Gemini 3 5 Flash Frontier Agents Coding Flash Tier Cost 2026 07 17 — Gemini 3.5 Flash multimodal analysis
• Mixture Of Experts Sparse Models Explained — MoE architecture deep-dive
• Agentic Coding Economics Roi Adoption 2026 05 18 — Agentic coding cost analysis