Inkling
by Thinking Machines Lab
Apache-2.0 975B-parameter open-weights model from Mira Murati's lab, built to be fine-tuned rather than rented
Inkling is Thinking Machines Lab's first open-weights foundation model, released 15 July 2026 under Apache 2.0. It is a 975-billion-parameter mixture-of-experts transformer with 41 billion active parameters and a 1-million-token context window. It exists for enterprises that need a US-developed open-weight model they can host on their own infrastructure and adapt to a domain, rather than renting a closed API they cannot inspect or customise.
Inkling is the first open-weights foundation model from Thinking Machines Lab, the public benefit corporation founded in February 2025 by former OpenAI CTO Mira Murati with John Schulman as chief scientist. Released on 15 July 2026 under the Apache 2.0 licence, it is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active per token, using 256 routed experts plus 2 shared experts per layer, a sigmoid router, interleaved sliding-window and global attention at a 5:1 ratio, relative positional embeddings in place of RoPE, and short convolutions throughout. It was pretrained on 45 trillion tokens of text, images, audio and video, and supports context windows up to 1 million tokens. A distilled Inkling-Small with 276B total and 12B active parameters followed on 31 July 2026. Reported benchmarks include 97.1% on AIME 2026, 87.2% on GPQA Diamond and 77.6% on SWE-Bench Verified. Artificial Analysis independently placed it at 41 on its Intelligence Index — the leading US open-weights release, ahead of Nemotron 3 Ultra at 38 and gpt-oss-120b at 24 — and measured notably better token efficiency, averaging 25K output tokens per index task against 43K for GLM-5.2 and 38K for Kimi K2.6. Thinking Machines does not meter API access to Inkling; it monetises through Tinker, the LoRA fine-tuning API launched in October 2025 whose customers include Bridgewater Associates. Weights ship on Hugging Face as original and NVFP4 checkpoints, with hosted inference from Together AI, Fireworks, Modal, Databricks and Baseten, and support in SGLang, vLLM, llama.cpp and Transformers.
A head of AI or ML platform lead at a Western regulated enterprise that is blocked from deploying Chinese open-weight models on procurement or policy grounds but still needs weights it can host and fine-tune in its own environment.
A US-developed, Apache-2.0 frontier-class base model that can be self-hosted and domain-adapted on proprietary data without per-token API fees or vendor lock-in.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Free, Usage-based, Contact for pricing
- Target Market
- CTOs, Heads of AI, Data Scientists, ML Platform Engineers, Enterprise Developers
- Deployment
- Open-source, Self-hosted, API-based, Multi-cloud
- Founded
- 2025
- Headquarters
- San Francisco, California, USA
- Team Size
- 51-200
- Customers
- 277,251 Hugging Face downloads in the month after release; Tinker customers include Bridgewater Associates
Key Features
- ✓Apache 2.0 open weights
Full weights are downloadable and commercially redistributable without a bespoke licence negotiation, unlike most frontier-lab community licences
- ✓975B mixture-of-experts architecture
256 routed plus 2 shared experts per layer with only 41B parameters active per token, keeping inference cost far below the total parameter count
- ✓1-million-token context window
Supports whole-repository or whole-corpus reasoning, with 64K and 256K options available on the Tinker platform for cheaper runs
- ✓Native multimodal pretraining
Trained on 45 trillion tokens spanning text, images, audio and video, scoring 73.5% on MMMU Pro and 91.4% on VoiceBench
- ✓Tinker LoRA fine-tuning API
Adapts the model on proprietary data through four primitives with infrastructure abstracted away, and Thinking Machines states customer data is never used to train its own models
- ✓Broad inference-stack support
Runs on SGLang, vLLM, llama.cpp and Hugging Face Transformers, with NVFP4 quantised checkpoints cutting VRAM needs from 2TB to about 600GB
- ✓High token efficiency
Averages 25K output tokens per Artificial Analysis Intelligence Index task versus 37-43K for GLM-5.2, Kimi K2.6 and DeepSeek v4 Pro, lowering real inference spend
Capabilities
Use Cases
- •Domain-adapted enterprise copilot
Fine-tune on internal documentation and ticket history so a support copilot answers in company-specific terms that a generic frontier API cannot match
- •Sovereign or air-gapped deployment
Run frontier-class inference entirely inside a controlled network boundary where sending prompts to a hosted API is not permitted
- •Cost conversion for high-volume inference
Organisations with committed GPU capacity replace per-token API spend with amortised infrastructure, aided by the model's low output-token counts
- •Agentic task automation
Inkling scores 1238 Elo on GDPval-AA v2, ahead of Kimi K2.6 at 1190 and DeepSeek v4 Flash max at 1189, supporting multi-step tool-using workflows
- •Derivative model development
Apache 2.0 terms let vendors and research teams build, brand and commercially redistribute their own fine-tuned variants without licence negotiation
Ideal For
Best For
- ✓Regulated Western enterprises barred by policy or procurement from deploying DeepSeek, Qwen, GLM or other Chinese open-weight models
- ✓Domain adaptation workloads — copilots, customer service, document understanding and agentic tasks — where fine-tuned specificity beats generic frontier performance
- ✓Organisations with existing GPU capacity that want to convert fixed infrastructure cost into inference rather than paying per-token API fees
- ✓Air-gapped or sovereign deployments where model weights must reside inside a controlled boundary
- ✓Research teams that need permissive Apache 2.0 terms for derivative models and commercial redistribution
Not Ideal For
- ✗Teams without serious GPU infrastructure — full-model deployment needs at least 2TB of aggregated VRAM (eight NVIDIA B300s or sixteen H200s), and even the quantised build needs 600GB; Pareekh Consulting's Pareekh Jain notes closed-model APIs stay more economical for many organisations
- ✗Buyers who need best-available raw capability, since Inkling trails DeepSeek V4 Pro and GLM 5.2 on SWE-Bench Verified and scores just +2 on Artificial Analysis's AA-Omniscience with a 63% hallucination rate
- ✗Medical, legal or safety-critical decision-making, which the model card explicitly rules out without additional fine-tuning, domain validation and human oversight
- ✗Teams expecting fine-tuning to preserve safety behaviour by default — analysts warn customisation can compromise safety filters, so alignment must be retested after every adaptation
Integrations
Deployment
Market & Ratings
277,251 Hugging Face downloads in the month after release; Tinker customers include Bridgewater Associates
Market Analysis
Pros
- ✓Apache 2.0 licensing with full weights on Hugging Face removes both licence negotiation and the vendor lock-in of a closed API
- ✓Independently verified as the leading US open-weights model on Artificial Analysis's Intelligence Index at 41, ahead of Nemotron 3 Ultra (38) and gpt-oss-120b (24)
- ✓Best-in-class token efficiency at 25K output tokens per index task, roughly 40% below the leading Chinese open models, which compounds into real inference savings
- ✓Strong agentic and multimodal results — 1238 Elo on GDPval-AA v2, 73.5% MMMU Pro, 91.4% VoiceBench — from a single model rather than a stack of specialists
- ✓Backed by $2B in Series A funding at a $12B valuation with Nvidia, AMD and Cisco as investors, so the lab is well capitalised to sustain the model line
Cons
- ✗Hardware floor is punishing: 2TB aggregated VRAM (eight B300s or sixteen H200s) for the full model, 600GB quantised, which Pareekh Consulting's Pareekh Jain says leaves closed APIs more economical for many organisations
- ✗Scores only +2 on Artificial Analysis's AA-Omniscience with 40% accuracy against a 63% hallucination rate, below leading open-weight peers on factual reliability
- ✗Trails DeepSeek V4 Pro and GLM 5.2 on SWE-Bench Verified (77.6%), so it is not the strongest option where raw coding capability is the deciding factor
- ✗The model card documents a tendency to comply with role-play and indirectly framed harmful prompts, requiring defence-in-depth moderation such as Llama Guard for consumer-facing use
- ✗Fine-tuning can compromise safety filters, so every domain adaptation needs its own safety retest rather than inheriting the base model's evaluation results
- ✗Thinking Machines is a young lab — founded February 2025, roughly 100 staff, and it lost Barret Zoph and Luke Metz back to OpenAI in January 2026 — so long-term model-line continuity is unproven
Pricing
Open weights (self-hosted)
$0
- ✓Apache 2.0 licence
- ✓Original and NVFP4 checkpoints on Hugging Face
- ✓Commercial use and redistribution permitted
- ✓No API metering by Thinking Machines
Tinker fine-tuning
Usage-based per million tokens
- ✓LoRA fine-tuning across 30+ models
- ✓Checkpoint storage at $0.10 per GB-month
- ✓Customer data never used to train Thinking Machines models
- ✓Downloadable checkpoints via API
Hosted inference via partners
From $1.87 per 1M input tokens
- ✓64K context: $1.87 input / $4.68 output per 1M
- ✓256K context: $3.74 input / $9.36 output per 1M
- ✓Available on Together AI, Fireworks, Modal, Databricks and Baseten
Thinking Machines deliberately does not meter API access to Inkling — the weights are Apache 2.0 and free — and monetises instead through Tinker, its LoRA fine-tuning API priced per million tokens with checkpoint storage at $0.10 per GB-month. Third-party hosted inference from Artificial Analysis's measurement runs $1.87 input / $4.68 output per million tokens at 64K context, doubling to $3.74 / $9.36 at 256K. The real cost of self-hosting is hardware, not licence: at least 2TB of aggregated VRAM for the full model or roughly 600GB quantised, which is why closed APIs remain cheaper below high sustained volume.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 4 on domains other than thinkingmachines.ai.
- 1.thinkingmachines.ai — introducing inklingvendor
- 2.thinkingmachines.ai — tinkervendor
- 3.huggingface.co — Inkling
- 4.artificialanalysis.ai — thinking machines has released inkling the new leading u s o
- 5.infoworld.com — thinking machines offers enterprises a us alternative in ope
- 6.en.wikipedia.org — Thinking Machines Lab
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Fundamental NEXUS
A foundation model built for tables, not text — enterprise prediction without feature engineering
Meta Muse Glimmer
Apache 2.0 open-weight 30B model built to run always-on AI agents on a single consumer GPU
Solar Pro 4
Agentic enterprise LLM tuned for multi-step document work at a tenth of frontier-model token cost
OpenRouter
One OpenAI-compatible API that routes every request across 500+ models and 80+ inference providers