AI Policy Wiki
Dashboard

Inkling Model Card (Thinking Machines Lab, July 2026)

high confidence · updated 2026-07-25

The model card for Thinking Machines Lab's first release: a 975B-total / 41B-active multimodal MoE under Apache 2.0 with a 1M-token context window. Carries the full architecture, self-hosting hardware requirements, comparative benchmark table against open- and closed-weight peers, and the safety conclusion that Inkling presents no material uplift beyond the existing open-weight ecosystem.

The Inkling model card is the formal release document for Thinking Machines Lab's first model, Inkling, published July 15, 2026. It is structured to the emerging model-card convention — general information, model properties, distribution, training, evaluations, safety, limitations, legal — and is the primary source for the model's architecture, hardware requirements, and safety case.

FieldValue
ProviderThinking Machines Lab, Inc.
ReleasedJuly 15, 2026
LicenseApache 2.0
Parameters975B total / 41B active
Context windowup to 1M tokens
Input modalitiesText, images, audio
Output modalitiesText only

Architecture

Inkling is described as a 66-layer decoder-only transformer with a sparse Mixture-of-Experts feed-forward backbone: each token is routed to 6 of 256 experts, plus 2 shared experts active on every token. Attention is a hybrid of local and global layers. The model is natively multimodal — images encoded via a hierarchical patch encoder, audio via discrete token encoding — with all modalities projected into a shared hidden space and processed jointly by the decoder. Supported numerics are BF16, MXFP8, and NVFP4.

Input constraints are specified: images at any pixel-based input with optimal dimensions 40px–4096px, and audio as 16kHz WAV with optimal length within 20 minutes.

Distribution and self-hosting

Available through Tinker, Thinking Machines' fine-tuning service, through third-party inference providers, and as downloadable weights on Hugging Face. The card states self-hosting requirements explicitly, which is the practically load-bearing detail for an open-weight release:

CheckpointMinimum VRAMExample configuration
BF16≥2 TB aggregated8× NVIDIA B300 or 16× NVIDIA H200
NVFP4 quantized≥600 GBW4A4 on 4× B300 (SM100+ required), or W4A16 on 8× H200

Supported inference frameworks: SGLang, vLLM, TokenSpeed, Unsloth, or Hugging Face.

Training data

The card describes data types as text, images, audio, and video, with provenance "drawn from publicly available sources, acquired from third parties, or synthetically generated or augmented," where publicly available data "includes content from the public internet and publicly accessible repositories." Curation is described as cleaning, processing, and modifying datasets, with deduplication and filtering "to remove junk or low-quality data, or to advance safety or other objectives." A separate training-data documentation page is referenced. See Training Data Walls.

Evaluations

All Inkling scores are reported at effort=0.99, against a comparison set of open-weight models (Nemotron 3 Ultra, Kimi K2.5, Kimi K2.6, GLM 5.2, DeepSeek V4 Pro) and closed-weight models (Gemini 3.1 Pro, Claude Fable 5, GPT 5.6 Sol). Selected rows:

BenchmarkInklingGLM 5.2DeepSeek V4 ProGemini 3.1 ProFable 5GPT 5.6 Sol
HLE (text only)29.7%40.1%35.9%44.7%53.3%47.2%
HLE (with tools)46.0%54.7%48.2%51.4%64.5%55.0%
AIME 202697.1%99.2%96.7%98.3%99.9%99.9%
GPQA Diamond87.2%89.5%88.8%94.1%92.6%94.1%
SWEBench Verified77.6%80.0%80.6%80.6%95.0%82.2%
SWEBench Pro (Public)54.3%62.1%55.4%54.2%80.0%64.6%
Terminal Bench 2.163.8%82.7%64%73.8%84.6%89.5%
GDPVal-AA v2 (Elo)12381514130796217601748
MCP Atlas74.1%77.8%73.2%78.2%83.3%81.8%
SimpleQA Verified43.9%38.1%57.0%77.3%68.3%71.6%
IFBench79.8%73.3%76.5%77.1%63.5%72.7%
FORTRESS (Adversarial)78.0%71.3%36.0%65.2%96.0%82.4%
FORTRESS (Benign)95.9%90.0%98.5%98.0%55.1%98.1%
StrongREJECT98.6%98.5%98.6%98.0%98.7%98.5%

The card also reports audio results — Audio MC 56.6%, MMAU 77.2%, VoiceBench 91.4% — against Gemini 3.1 Pro at 66.8%, 82.5%, and 94.3% respectively; the other comparators were not evaluated on audio.

The FORTRESS pair is worth reading together: Fable 5 scores highest on the adversarial split (96.0%) and lowest by a wide margin on the benign split (55.1%), the signature of a strict refusal posture, while Inkling sits at 78.0% and 95.9%. The card discloses two methodology notes: Terminal Bench 2.1 rollouts with solution contamination from web search were scored 0, and Fable 5 and GPT 5.6 Sol were benchmarked on CharXiv RQ (with python) using Thinking Machines' internal harness. See AI Benchmarks and Evaluation.

Safety

Safety evaluation is reported across three areas, conducted before release with attention to whether safety behavior held consistently across text, audio, and image inputs.

  • Everyday interaction — sycophancy, harmful manipulation, and psychological-harm patterns including parasocial dependency and validation of delusional reasoning, tested partly through multi-turn open-ended external red-teaming "designed to surface issues that only emerge over longer conversations." Refusal of genuinely harmful requests was assessed alongside over-refusal of benign ones. See Companion Chatbot Harms — Cross-Cutting Analysis.
  • CBRN and cyber — knowledge and procedural uplift assessed through internal evaluations, external testing, and refusal-suppressed variants "intended to estimate latent capability with safeguards removed." See CBRN Uplift.
  • Loss of control — agentic capability, strategic deception, and sabotage potential benchmarked against public frontier models, with the model found "materially below frontier capabilities."

The stated conclusion is the open-weight-specific one: "Across all areas, we concluded that Inkling did not present risk of material uplift beyond what's already available in the open-weight ecosystem." This is a relative rather than absolute safety claim, benchmarked against what an adversary could already obtain.

Residual risk is disclosed: an "occasional tendency to comply with role-play and indirectly framed prompts concerning harmful topics," described as consistent with other open-weight models and best addressed by defense in depth — downstream moderation tools such as Llama Guard layered around the model — "rather than relying on model refusals alone." See Open-Weight Frontier Models.

Limitations

The card lists hallucination, occasional instruction-following failures, degraded performance in long multi-turn conversations, potential demographic, cultural, or linguistic bias from web-derived and synthetic training data, uneven performance across languages and less-represented domains, and a training cutoff. Downstream developers are advised to run their own use-case evaluations, implement application-layer safeguards (content filtering, rate limiting, monitoring), and avoid deployment in medical, legal, or safety-critical decision-making without additional fine-tuning, domain validation, and human oversight.

Provenance

Fetched from thinkingmachines.ai, the developer's canonical domain, and linked from the official announcement post. Specifications (975B/41B, 1M context, Apache 2.0, July 15 2026 release) corroborated against the announcement, the Hugging Face repository, and independent coverage from MarkTechPost and TechCrunch dated July 15, 2026.

Relationships