Inkling is the first model released by Thinking Machines Lab, published with full open weights under the Apache 2.0 license on July 15, 2026. It is a natively multimodal mixture-of-experts transformer with 975 billion total parameters, of which roughly 41 billion are active per token, and a 1-million-token context window (Source: thinkingmachines.ai; thinkingmachines.ai). The company positions the model for enterprise fine-tuning through its Tinker platform rather than as a frontier-capability leader, stating at release that Inkling is "not the strongest overall model available today, open or closed" (Source: techcrunch.com). Press coverage described it as the largest US-built open-weight model publicly available at release (Source: techtimes.com).
| Field | Value | |
|---|---|---|
| Developer | [[companies/thinking-machines-lab | Thinking Machines Lab]] |
| Released | July 15, 2026 | |
| License | Apache 2.0 | |
| Architecture | 66-layer decoder-only transformer, sparse mixture-of-experts | |
| Parameters | 975B total / ~41B active per token (6 of 256 routed experts + 2 shared) | |
| Context window | 1 million tokens | |
| Modalities | Text, image, and audio input; text output | |
| Training data | 45 trillion tokens of text, image, audio, and video | |
| Training hardware | Nvidia GB300 NVL72 systems | |
| Weights | Open, published on Hugging Face (BF16 and NVFP4 checkpoints) | |
| Companion release | Inkling-Small preview (276B total / 12B active) |
Capabilities and benchmarks
Thinking Machines trained Inkling as a broad generalist across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks, rather than optimizing for a single domain, describing breadth as the property that matters for customization (Source: thinkingmachines.ai).
Self-reported benchmark scores at release (July 15, 2026; all Inkling evaluations run at effort 0.99 and temperature 1.0, with scores for several evaluations taken from Artificial Analysis for both Inkling and external models), against selected open- and closed-weight peers:
| Benchmark | Inkling | [[models/kimi-k2\ | Kimi K2.6]] | [[models/glm-5-2\ | GLM 5.2]] | [[models/deepseek-v4\ | DeepSeek V4 Pro]] | [[models/claude-fable-5\ | Claude Fable 5]] (max) | [[models/gpt-56\ | GPT 5.6 Sol]] (max/xhigh) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| HLE (text only) | 29.7% | 35.9% | 40.1% | 35.9% | 53.3% | 47.2% | |||||
| HLE (with tools) | 46.0% | 54.0% | 54.7% | 48.2% | 64.5% | 55.0% | |||||
| AIME 2026 | 97.1% | 96.4% | 99.2% | 96.7% | 99.9% | 99.9% | |||||
| GPQA Diamond | 87.2% | 91.1% | 89.5% | 88.8% | 92.6% | 94.1% | |||||
| SWE-bench Verified | 77.6% | 80.2% | 80.0% | 80.6% | 95.0% | 82.2% | |||||
| SWE-bench Pro (public) | 54.3% | 58.6% | 62.1% | 55.4% | 80.0% | 64.6% | |||||
| Terminal Bench 2.1 | 63.8% | 71.3% | 82.7% | 64% | 84.6% | 89.5% | |||||
| GDPVal-AA v2 (Elo) | 1238 | 1190 | 1514 | 1307 | 1760 | 1748 | |||||
| MCP Atlas | 74.1% | 68.1% | 77.8% | 73.2% | 83.3% | 81.8% | |||||
| BrowseComp (w/ ctx mgmt) | 77.1% | 83.2% | – | 83.4% | 88.0% | 90.84% | |||||
| SimpleQA Verified | 43.9% | 38.7% | 38.1% | 57.0% | 68.3% | 71.6% | |||||
| IFBench | 79.8% | 76.0% | 73.3% | 76.5% | 63.5% | 72.7% | |||||
| MMMU Pro (Standard 10) | 73.5% | 79.0% | – | – | 84.2% | 83.0% | |||||
| CharXiv RQ (with python) | 82.0% | 86.7% | – | – | 89.4% | 87.8% | |||||
| VoiceBench | 91.4% | – | – | – | – | – | |||||
| FORTRESS (Adversarial) | 78.0% | 65.6% | 71.3% | 36.0% | 96.0% | 82.4% | |||||
| StrongREJECT | 98.6% | 99.8% | 98.5% | 98.6% | 98.7% | 98.5% |
(Source: thinkingmachines.ai; thinkingmachines.ai. Terminal Bench 2.1 figures for Inkling use an internal coding harness, with contaminated rollouts scored 0; external models' scores are self-reported where available.)
The pattern across the table: Inkling trails the strongest Chinese open-weight models (GLM 5.2, Kimi K2.6) on several coding and reasoning benchmarks while leading the compared open-weight group on instruction following (IFBench), short-form factuality (SimpleQA Verified), audio, and adversarial-refusal safety (FORTRESS), consistent with the company's positioning of the model as a balanced customization base rather than a benchmark leader (Source: marktechpost.com). On Design Arena's Agentic Web Dev leaderboard, a blinded human evaluation of generated web apps, Inkling scored 1257 at release, among the strongest open-weight entries (Source: thinkingmachines.ai).
Controllable thinking effort
Inkling exposes a continuous thinking-effort setting (swept from 0.2 to 0.99 in the release evaluations) that trades performance against generated tokens; in Hugging Face transformers it is exposed as a reasoning_effort argument with named levels. Thinking Machines reports that Inkling matches Nvidia's Nemotron 3 Ultra on Terminal Bench 2.1 at roughly one third of the tokens (Source: thinkingmachines.ai; marktechpost.com).
Calibration and epistemics
The company trained the model for calibration, instruction following, and what it calls resistance to censorship. On ForecastBench (Brier index, higher better) Inkling scored 61.1 without search and 63.7 with search, against 54.6/58.6 for Claude Opus 4.8 and 59.1/64.7 for GPT-5.5; on Prophet Arena it scored a Brier score of 0.1617 (results obtained June 30–July 13, 2026 on a different checkpoint than the one released). Calibration was trained via RL against proper scoring rules on resolved real-world questions, plus abstention-aware short-form QA rewards; instruction following used paired rubric and claim-verification graders. Cognition evaluated the model on its Propaganda and Censorship Eval and reported strong censorship non-compliance (Source: thinkingmachines.ai).
Training and architecture
Inkling is a 66-layer decoder-only transformer with a sparse mixture-of-experts feed-forward backbone: each token is routed to 6 of 256 experts, with 2 shared experts active on every token, using a sigmoid-based router with an auxiliary-loss-free load-balancing bias; the MoE design largely follows DeepSeek-V3. Attention interleaves sliding-window and global layers at a 5:1 ratio with 8 KV heads, encodes position with a relative positional embedding rather than RoPE (which the company reports extrapolates better to long sequences), and applies short convolutions after the key/value projections and on residual branch outputs (Source: thinkingmachines.ai; thinkingmachines.ai).
Multimodality is encoder-free: audio enters as dMel spectrograms and images as 40×40-pixel patches through a four-layer hMLP, both projected via a lightweight embedding layer and processed jointly with text tokens — a design choice tied to the model's intended role as the background reasoning model in the company's interaction-models system (Source: thinkingmachines.ai).
Inkling was pretrained from scratch on 45 trillion multimodal tokens (text, image, audio, and video) on Nvidia GB300 NVL72 systems, using a hybrid optimization strategy — Muon for large matrix weights, Adam for other parameters — with weight-decay strength coupled to the square of the learning rate (Source: thinkingmachines.ai; axios.com). Post-training was bootstrapped with an initial SFT pass on synthetic data generated by other open-weight models, including Moonshot AI's Kimi K2.5; the company states the bootstrap accounts for a small fraction of compute (Source: techcrunch.com; thinkingmachines.ai). The use of a Chinese open-weight model in an American lab's bootstrap pipeline is an instance of the cross-border reuse dynamics described at Open-Weight Frontier Models and Distillation.
The majority of post-training compute went to large-scale asynchronous reinforcement learning, scaled past 30 million rollouts across two long continuous runs, with reward on a held-out aggregate of reasoning evaluations improving log-linearly throughout (Source: thinkingmachines.ai; axios.com). Thinking Machines did not disclose Inkling's training cost; Emad Mostaque estimated its pretraining at $10–20 million at GB300 rental rates (Source: thinkingmachines.ai). Effort controllability was trained by varying the system message and per-token cost across rollouts. The company also reports an emergent, untargeted compression of the model's chain of thought over RL training — reasoning traces dropped grammatical overhead while remaining comprehensible, an effect it notes Cognition also observed while training SWE-1.7 (Source: thinkingmachines.ai).
Safety and evaluations
Thinking Machines published a model card at release documenting pre-release safety evaluations across everyday human-AI interaction (sycophancy, harmful manipulation, parasocial dependency, validation of delusional reasoning, including multi-turn external red-teaming), CBRN and cyber uplift (internal evaluations, external testers, and refusal-suppressed variants to estimate latent capability with safeguards removed), and loss of control (agentic capability, strategic deception, and sabotage potential, which it found materially below frontier capabilities). The company concluded that Inkling "did not present risk of material uplift beyond what's already available in the open-weight ecosystem" (Source: thinkingmachines.ai).
On FORTRESS, a benchmark testing refusal of weapons- and violence-related requests alongside benign look-alikes, Inkling posted the strongest adversarial score (78.0%) of the open-weight models the company compared, without over-refusing benign analogs (95.9% benign); it scored 98.6% on StrongREJECT (Source: thinkingmachines.ai). The model card identifies an occasional tendency to comply with role-play and indirectly framed prompts on harmful topics as residual risk, which it describes as consistent with other open-weight models and best addressed with downstream defense-in-depth such as input/output classifiers (Source: thinkingmachines.ai).
In an August 2026 study probing whether open-weight models respond unusually strongly to decoded proprietary reasoning, Inkling was used as the principal control and largely behaved as one. Across 90 problems, a four-word GPT-5.6 Sol prefill left it at classifier AUC 0.52 against its own control, though a Claude Opus 4.8 prefill moved it to 0.70; on a 30-problem Humanity's Last Exam set it showed no shift in visible-answer n-gram overlap toward Opus where Kimi-K3 shifted on 29 of 30 problems, and it required more than 10^14 queries to reproduce a 16-token span of an Opus answer. The authors state the analysis "cannot causally establish distillation" in either direction (Stealing Reasoning Traces from Proprietary LLM APIs).
Thinking Machines has not issued a public safety policy comparable to other frontier labs' responsible-scaling frameworks (Responsible Scaling Policy (RSP)), and no third-party government-institute evaluation (such as by the UK or US AISI) had been published as of July 16, 2026.
Availability and pricing
Full weights are available on Hugging Face in the original BF16 checkpoint and an NVFP4 checkpoint for Nvidia Blackwell systems; self-hosting the BF16 checkpoint requires at least 2 TB of aggregated VRAM (8× Nvidia B300 or 16× H200), while NVFP4 reduces the requirement to about 600 GB (Source: thinkingmachines.ai). Fine-tuning is offered through the company's Tinker API with 64K and 256K context-length options, at a 50% limited-time discount, alongside an Inkling Playground chat interface in the Tinker console (Source: thinkingmachines.ai).
Hosted inference APIs launched day-one via TogetherAI, Fireworks, Modal, Databricks, and Baseten, with open-source inference support in SGLang, vLLM, TokenSpeed, llama.cpp (via Unsloth), and Hugging Face transformers; a multi-token-prediction drafter ships for speculative decoding (Source: thinkingmachines.ai; marktechpost.com).
The release extends the company's stated "customizable, understandable" strategy: rather than competing on peak capability, Inkling is presented as a base for enterprise-specific fine-tunes, consistent with the Tinker financial-services results reported with Bridgewater's AIA Labs in June 2026 (see Thinking Machines Lab).
Reception
Coverage at release centered on the model's deliberate positioning below the capability frontier. TechCrunch framed the release as Thinking Machines "amp[ing] up its bet against one-size-fits-all AI," noting the company's own statement that stronger models exist both open and closed (Source: techcrunch.com). MarkTechPost's technical assessment credited the Apache 2.0 licensing, day-zero ecosystem support, token-efficiency curves, and FORTRESS results, while noting that Inkling trails GLM 5.2 and Kimi K2.6 on HLE, Terminal Bench 2.1, and SWE-bench Verified, that its Terminal Bench numbers use an internal harness unlike self-reported competitor scores, and that Inkling-Small's weights were unreleased at launch (Source: marktechpost.com). The Kimi K2.5 bootstrap disclosure drew attention in the context of US-China model-reuse debates (AI Race Dynamics); Fortune noted the disclosure in its coverage of Moonshot AI's Kimi K3 release the following day (Source: fortune.com).
Related models
Inkling-Small, a 276-billion-parameter mixture-of-experts model with 12B active parameters, was shared as a preview alongside the release (Source: axios.com). Trained with a similar recipe and the same post-training stack, the preview checkpoint matched or exceeded the larger model on many benchmarks — including HLE with tools (46.6% vs. 46.0%), GPQA Diamond (88.3% vs. 87.2%), and IFBench (83.4% vs. 79.8%) — which the company attributes to pre-training data and recipe improvements made for the smaller model, while trailing on Terminal Bench 2.1 (52.7% vs. 63.8%), Tau 3 Banking (13.6% vs. 23.7%), and SimpleQA Verified (20.9% vs. 43.9%). These are preview-checkpoint figures; the released model's differ, and are recorded on Inkling-Small. Inkling-Small was released on July 30, 2026 as an open-weights mixture-of-experts transformer with 276 billion total and 12 billion active parameters, trained on NVIDIA GB300 NVL72 systems, with native reasoning over audio and images and a context window of up to 1 million tokens. Thinking Machines said it reaches performance comparable to Inkling — 975 billion total, 41 billion active — at a quarter of the size (Source: thinkingmachines.ai). The company describes Inkling as the first in a family of models of different sizes.
Relationships
- instance-of: Open-Weight Frontier Models
- depends-on: Kimi K2 (synthetic post-training data in early bootstrap)
- related: Thinking Machines Lab, Distillation, AI Race Dynamics, DeepSeek-V3 (MoE design lineage), Kimi K3 (released the following day; compared open-weight flagship)