Inkling-NVFP4
The first model from Thinking Machines — a 975B parameter Mixture of Experts trained on 45T+ tokens of text, image, and audio, with strong performance on agentic and reasoning-heavy tasks.
Total Parameters
975B
41B Activated
Context Window
977K
Tokens
Modalities
Text, Image
& Audio
Released
Jul 2026
Thinking Machines
Thinking Machines' first frontier release
Inkling is the debut model from Thinking Machines, a large Mixture of Experts architecture with 975B total parameters and 41B active at inference. Pretrained on more than 45 trillion tokens of text, image, and audio data, it supports a native 977K context length and offers strong performance across agentic and reasoning-heavy workloads — from long-context analysis to tool-use pipelines and multi-step coding tasks.
Built for agentic and reasoning workloads
Agentic Workflows
Strong tool-use and multi-turn planning make Inkling well-suited for long-horizon agents that reason across many steps.
Reasoning-Heavy Tasks
41B active parameters trained on 45T+ tokens deliver frontier-level reasoning across math, science, and coding benchmarks.
Long-Context Analysis
A native 977K context window lets you process entire codebases, contract libraries, or research corpora in a single call.
Multimodal Understanding
Text, image, and audio pretraining data enables broad multimodal comprehension for document, transcript, and visual pipelines.
Artificial Analysis Metrics
Frontier-level reasoning with strong coding and agentic scores.
Intelligence Index
Better than 78% of models
GPQA Diamond
Better than 82% of models
MMLU-Pro
Better than 80% of models
| Category | Benchmark | Score | Description |
|---|---|---|---|
| Reasoning | Intelligence Index | 41 | Artificial Analysis composite score |
| Reasoning | GPQA Diamond | 79% | Graduate-level scientific reasoning |
| Reasoning | MMLU-Pro | 82% | Multitask language understanding (pro) |
| Reasoning | Humanity's Last Exam | 18% | Expert-level cross-domain reasoning |
| Reasoning | AA-LCR | 61% | Long-context reasoning evaluation |
| Reasoning | IFBench | 68% | Instruction-following accuracy |
| Coding | LiveCodeBench | 63% | Competitive programming problems |
| Coding | SciCode | 42% | Python for scientific computing |
| Coding | Terminal-Bench Hard | 22% | Agentic coding & terminal use |
| Agentic | τ²-Bench Telecom | 58% | Agents in dual-control scenarios |
| Knowledge | AA-Omniscience | 34% | Broad-domain factual accuracy |
Metrics sourced from Artificial Analysis.
Flexible Pricing Tiers
Choose the optimal balance of speed and cost. Prices are per 1M tokens.
| Tier | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Batch | $0.60 | $2.00 |
| Async | $0.90 | $3.00 |
| Realtime | $1.20 | $4.00 |
Context window natively supported up to 977k tokens.
Start Building in Minutes
Inkling-NVFP4 is accessible via OpenAI-compatible endpoints. Here's how to call it with the standard Python SDK via Doubleword.
Tip: Route long-running jobs to Async or Batch
For evals, agent pipelines, and offline data generation, using the Async or Batch tiers cuts cost by 25–50% vs the Realtime tier at identical model quality.
from openai import OpenAI
client = OpenAI(
api_key="your-api-key-here",
base_url="https://api.doubleword.ai/v1"
)
response = client.chat.completions.create(
model="thinkingmachines/Inkling-NVFP4",
messages=[
{"role": "user", "content": "Explain the tradeoffs of MoE architectures."}
],
)
print(response.choices[0].message.content)