Thinking Machines Lab, the AI research company founded by former OpenAI researchers, has released Inkling, a new open-weights Mixture-of-Experts model that developers can fine-tune and deploy today. With 975 billion total parameters, 41 billion active, a 1 million token context window, and native multimodal reasoning across text, images, and audio, Inkling represents a significant step toward practical, customizable foundation models for real-world applications.

Background: The Customization Gap

Open-weights models have democratized access to LLM capabilities, but a persistent gap remains: most releases are generalist checkpoints that require substantial fine-tuning compute to adapt to specialized domains. Meanwhile, closed models like GPT-5.6 and Claude Fable 5 offer powerful customization APIs but lock users into proprietary ecosystems with opaque pricing and data policies.

Thinking Machines Lab positions Inkling differently. The company's mission — building AI that extends human will and judgment — translates into a platform approach: release the full model weights, provide a fine-tuning platform (Tinker) with managed infrastructure, and partner across the inference ecosystem (Together AI, Fireworks, Modal, Databricks, Baseten, vLLM, SGLang, llama.cpp) so developers can deploy customized checkpoints wherever they choose.

Technical Architecture: MoE at Scale with Controllable Compute

Inkling follows a DeepSeek-V3-style Mixture-of-Experts design with deliberate departures for efficiency and long-context performance:

Expert Configuration: Each MoE layer contains 256 routed experts and 2 shared experts, with 6 routed experts active per token. A sigmoid-based router with auxiliary-loss-free load-balancing bias replaces the traditional top-k gating, reducing routing overhead while maintaining expert specialization.

Attention Design: Sliding-window and global attention layers interleave at a 5:1 ratio with 8 KV heads. Relative positional embeddings (Shaw et al., 2018; Huang et al., 2018) replace Rotary Positional Embeddings (RoPE) — the team reports better extrapolation to 1M token contexts with this choice. Short convolutions after key/value projections and on residual branch outputs further stabilize long-sequence training.

Multimodal Encoders: Audio inputs enter as dMel spectrograms (He Bai et al., 2024) via a lightweight embedding layer; images are encoded as 40×40 pixel patches through a 4-layer hMLP (Touvron et al., 2022). Both modalities project into the shared token space without separate encoders, enabling joint reasoning across text, vision, and audio in a single forward pass.

Training Scale: 45 trillion tokens spanning text, images, audio, and video. A hybrid optimizer — Muon for large matrix weights, Adam for others — with learning-rate-coupled weight decay (Kosson et al., 2023; Defazio, 2025) maintained weight norm stability across the training horizon.

Controllable Thinking Effort: The Developer Lever

The standout feature for practitioners is controllable thinking effort — a single scalar (0.2 to 0.99) passed via system message that adjusts the model's token budget for internal reasoning. At low effort, Inkling produces concise responses with minimal chain-of-thought; at high effort, it generates extended reasoning traces that improve performance on coding, math, and complex instruction following.

Benchmarks demonstrate the leverage: Inkling at effort=0.99 matches Nemotron 3 Ultra on Terminal Bench 2.1 (agentic coding) at roughly one-third the token cost. On Humanity's Last Exam (HLE), effort sweeps from 0.2 to 0.99 yield a 20+ percentage point accuracy gain. This is not a binary "thinking mode" toggle — it is a continuous compute dial that developers can tune per use case, per request, or even per sub-task within an agentic workflow.

Multimodal Capabilities: Audio and Vision as First-Class Citizens

Inkling was trained natively on audio and vision, not adapted post-hoc. The results are competitive with specialist omni models:

Benchmark	Inkling (effort=0.99)	Qwen3-Omni	Nemotron-3 Nano-Omni	Gemini 3.1 Pro (high)
VoiceBench	91.4%	88.8%	89.4%	94.3%
MMAU (Audio)	77.2%	77.5%	76.7%	82.5%
Audio MC	56.6%	24.3%	23.2%	66.8%
MMMU Pro	73.5%	75.0%	53.0%	82.0%
CharXiv RQ + Python	82.0%	78.7%	86.7%	89.9%

VoiceBench at 91.4% places Inkling among the strongest open-weights audio models. The Python tool integration for vision tasks (zooming, cropping, visual reasoning with code-based analysis) enables workflows where the model writes and executes code to process images — a pattern familiar to developers using agentic coding tools.

Agentic Coding and Tool Use: Built for Harnesses

Inkling was trained inside multiple coding and agent harnesses (OpenCode, Terminal Bench, SWE-Bench) with randomized tool schemas to reduce harness-specific overfitting. Results on agentic benchmarks:

Benchmark	Inkling	Nemotron 3 Ultra	Kimi K2.6	GLM 5.2	DeepSeek V4 Pro
SWE-Bench Verified	77.6%	70.7%	80.2%	80.0%	80.6%
Terminal Bench 2.1	63.8%	56.4%	71.3%	82.7%	64.0%
MCP Atlas	74.1%	44.7%	68.1%	77.8%	73.2%
Tau 3 Banking	23.7%	13.8%	20.6%	26.8%	25.8%

The SWE-Bench Verified score of 77.6% (bash-only harness) is competitive with leading closed models. More importantly, the model's controllable effort setting means teams can trade latency for accuracy on a per-task basis — critical for CI/CD pipelines where marginal test-time compute can flip a failing patch to passing.

Epistemics: Calibration, Instruction Following, and Censorship Resistance

Thinking Machines Lab emphasizes three pillars of model "epistemics":

1. Calibration: RL against proper scoring rules on resolved real-world questions. ForecastBench Brier Index of 61.1 (no search) and 63.7 (with search) exceeds GPT-5.5 and Claude Opus 4.8.
2. Instruction Following: Dual automated graders — a rubric grader for recall and a claims grader with agentic web search for factual verification — reduce hallucination without sacrificing helpfulness.
3. Censorship Non-Compliance: Evaluated on Cognition's Propaganda and Censorship Eval, Inkling exhibits strong patterns of refusing to comply with censorship directives while maintaining safety boundaries on genuinely harmful content.

For developers building applications where trust and verifiability matter — financial analysis, medical summarization, legal review — these properties translate to fewer silent failures and more reliable abstention when the model lacks confidence.

Safety: Strongest Open-Weights Refusals on FORTRESS

Model	FORTRESS Adversarial	FORTRESS Benign	StrongREJECT
Inkling (0.99)	78.0%	95.9%	98.6%
Nemotron 3 Ultra	77.6%	90.5%	98.7%
Kimi K2.5	54.1%	98.3%	99.5%
DeepSeek V4 Pro	36.0%	98.5%	98.6%

Inkling achieves the highest adversarial refusal rate among open-weights models on FORTRESS (weapons, violence, benign look-alikes) while maintaining 95.9% benign pass-through — avoiding the over-refusal that plagues many safety-tuned models.

Inkling-Small: The 12B Active Parameter Companion

Alongside the flagship model, Thinking Machines Lab released a preview of Inkling-Small (276B total, 12B active). Early benchmarks show it matching or exceeding the 41B active model on HLE (29.6% vs 29.7%), AIME 2026 (95.1% vs 97.1%), and SWE-Bench Verified (77.4% vs 77.6%). With 70% fewer active parameters, it targets workloads where latency and cost dominate: code grading, synthetic data generation, and high-throughput agentic loops.

Full weights will be released once testing completes.

Developer Experience: Fine-Tuning on Tinker and Deployment Anywhere

Tinker Platform: Inkling is available for fine-tuning on Tinker today with 64K and 256K context options at a 50% launch discount. Three new cookbook recipes showcase audio fine-tuning workflows. The tml-renderer Python package handles tool calls, reasoning traces, and multimodal inputs for post-training pipelines.

Inference Partners: Checkpoints deploy via Together AI, Fireworks, Modal, Databricks, Baseten APIs. Open-source inference support includes SGLang, Miles, vLLM (via Inferact), TokenSpeed (via Lightseek), llama.cpp (via Unsloth), and Hugging Face Transformers. The NVFP4 quantized checkpoint on Hugging Face enables efficient Blackwell inference.

Playground: The Inkling Playground in the Tinker console offers a chat interface with integrated agentic web search, free for a limited time — useful for qualitative evaluation before committing compute.

Comparison: Inkling in the Open-Weights Landscape

Feature	Inkling	Nemotron 3 Ultra	Kimi K2.6	DeepSeek V4 Pro	Qwen 3.5 Omni
Architecture	MoE (256 experts)	Dense	MoE	MoE	MoE
Active Params	41B	53B	32B	37B	~40B
Context Window	1M tokens	128K	256K	128K	1M
Modalities	Text, Image, Audio	Text	Text, Image	Text	Text, Image, Audio
Controllable Effort	Yes (0.2-0.99)	No	No	No	No
Fine-Tuning Platform	Tinker (managed)	Self-serve	Self-serve	Self-serve	Self-serve
Weights License	Apache 2.0	NVIDIA proprietary	Apache 2.0	Custom	Apache 2.0
SWE-Bench Verified	77.6%	70.7%	80.2%	80.6%	—
VoiceBench	91.4%	—	—	—	—
FORTRESS Adv.	78.0%	77.6%	54.1%	36.0%	—

Inkling's differentiator is the combination of controllable compute, native multimodality, and a managed fine-tuning path — not raw benchmark supremacy.

Practical Implications for Developers

1. Cost-Quality Tradeoff at Inference Time: Instead of choosing between a small fast model and a large accurate one, teams can deploy a single Inkling endpoint and dial effort per request. Low-effort for classification and extraction; high-effort for architectural reasoning and complex refactoring.

2. Audio-First Applications: VoiceBench 91.4% enables production voice assistants, meeting summarization with speaker diarization, and audio content moderation without stitching separate ASR + LLM + TTS pipelines.

3. Vision+Code Workflows: The Python tool integration for image reasoning (zoom, crop, analyze) mirrors how human developers debug visual issues — write a script, inspect, iterate. This is a natural fit for agentic UI testing, diagram-to-code generation, and technical documentation automation.

4. Forecasting and Decision Support: Calibrated probability outputs (ForecastBench Brier Index 61.1) make Inkling suitable for risk assessment, demand forecasting, and strategic planning where confidence intervals matter.

5. Data Sovereignty: Apache 2.0 weights + self-hosted inference options (vLLM, SGLang, llama.cpp) mean regulated industries can fine-tune and deploy without data leaving their VPC.

Frequently Asked Questions

How does controllable thinking effort work technically? During RL post-training, the system message specifies an effort level (0.2-0.99) and a per-token cost penalty. The model learns to emit longer or shorter reasoning traces to maximize reward minus cost. At inference, the same system message controls the behavior — no architecture changes required.

Can I fine-tune Inkling without Tinker? Yes. Full weights are on Hugging Face (Apache 2.0). The tml-renderer package and standard LoRA/QLoRA workflows work with any training stack. Tinker provides managed infrastructure, hyperparameter optimization, and evaluation — convenient but not required.

What are the hardware requirements for self-hosting? The NVFP4 quantized checkpoint runs on NVIDIA Blackwell (GB200/GB300) with ~200 GB VRAM for 41B active params. For Hopper/Ampere, standard FP8/INT4 quantization via llama.cpp or vLLM reduces VRAM to ~80-100 GB. Inkling-Small (12B active) fits on a single H100 or dual A100s.

Is the 1M token context actually usable? Benchmarks show meaningful performance at 256K+ tokens on HLE with tools and BrowseComp with context management. The relative positional embedding and sliding-window attention were designed for this regime. However, KV cache memory scales linearly — plan infrastructure accordingly.

How does Inkling compare to GPT-5.6 Sol or Claude Fable 5 on coding? Closed models still lead on SWE-Bench Verified (82-95% vs Inkling's 77.6%). The gap narrows at high effort settings and with domain-specific fine-tuning. For teams that can invest in customization, Inkling's open weights + Tinker fine-tuning can close the delta for specialized codebases.

What safety guarantees exist for open-weights deployment? FORTRESS adversarial refusal at 78% and StrongREJECT at 98.6% indicate strong built-in safeguards. However, open weights can be further fine-tuned to remove safeguards. Thinking Machines Lab publishes safety evals transparently and monitors misuse; enterprises should run their own red-teaming before production deployment.

Key Takeaways

• Inkling is a 975B/41B MoE model with native text, image, and audio reasoning — the first open-weights release from Thinking Machines Lab (ex-OpenAI researchers).
• Controllable thinking effort (0.2-0.99) lets developers trade latency for accuracy per request — matching larger models at lower token cost on coding and reasoning benchmarks.
• Fine-tuning is available today on Tinker (managed) or self-serve via Hugging Face weights — with deployment partners covering Together AI, Fireworks, vLLM, SGLang, llama.cpp, and more.
• Audio performance (VoiceBench 91.4%) and vision+code tool use enable new agentic workflows — voice assistants, multimodal coding agents, technical diagram reasoning.
• Apache 2.0 license and self-hosted inference options — suitable for regulated environments and data sovereignty requirements.
• Inkling-Small (12B active) preview matches the flagship on many benchmarks — a cost-optimized sibling for high-throughput workloads.

Conclusion

Inkling does not claim to be the strongest model on every benchmark. It claims to be the most customizable foundation model for developers who need multimodal reasoning, compute flexibility, and deployment freedom. By releasing full weights, providing a managed fine-tuning platform, and partnering across the inference ecosystem, Thinking Machines Lab is betting that the next wave of AI adoption will be driven not by benchmark leaderboards but by how easily teams can mold a model to their specific problem.

For developers evaluating open-weights models in July 2026, Inkling deserves a spot on the shortlist — particularly for audio-first applications, agentic coding with variable compute budgets, and teams that need to fine-tune on proprietary data without leaving their infrastructure.

---

Source: Thinking Machines Lab - Introducing Inkling