Model
Model explorer

Inkling

OPEN
Thinking Machines Lab · Inkling family · released Jul 15, 2026

Thinking Machines Lab's (Mira Murati's startup) first public model release — a 975B/41B-active MoE with adjustable inference-time "thinking effort", calibrated uncertainty, and text/vision/audio input.

ReasoningCodingVisionFunction callingTool useAgentic
2194.1
Elo · rank #70
Parameters
975B
Active params
41B (MoE)
Context
1M tokens
Architecture
Sparse MoE, 66 layers (975B total / 41B active, 6-of-256 routed experts + 2 shared experts/token), hybrid sliding-window/global attention (5:1), relative positional features (no RoPE)
License
Apache 2.0
Languages
—
API price (in/out)
$1.87 / $4.68
Modalities
text · vision · audio
Benchmark results
Bar shows position within the tracked field; marker = field best
AIMEMath97.1%#7
best: GPT-5.2 · 100.0%
ARC-AGI-1Reasoning79.5%#25
best: Claude Fable 5 · 98.5%
ARC-AGI-2Reasoning36.5%#32
best: GPT-6 Astra · 95.0%
BrowseCompAgents77.1%#22
best: Kimi K3 · 91.2%
CharXivVision78.1%#22
best: Qwen3.8-Flash-Next · 90.6%
GDPval-AAAgents1238#35
best: Claude Fable 5 · 1932
GPQA DiamondReasoning87.2%#52
best: GPT-6 Astra · 96.0%
Humanity's Last ExamReasoning29.7%#58
best: Claude Opus 5.5 · 67.7%
IFBenchReasoning79.8%#9
best: MiniMax M3 · 83.0%
MCP AtlasAgents74.1%#6
best: Step 5 Preview · 85.6%
MMMU-ProVision73.5%#36
best: Claude Opus 4.7 · 85.5%
SimpleQAKnowledge43.9%#7
best: GPT-4.5 · 62.5%
SWE-bench ProCoding54.3%#40
best: Claude Opus 5.5 · 89.9%
SWE-bench VerifiedCoding77.6%#40
best: Claude Opus 5 · 97.0%
Terminal-Bench 2.0Coding63.8%#42
best: DeepSeek-V4.1-Flash · 90.6%
Run it locally
VRAM @ Q4
600 GB
VRAM @ FP16
1900 GB
Fits on (Q4)
Multi-node cluster required
Official NVFP4 checkpoint needs Blackwell-class GPUs (~600GB VRAM); community Unsloth GGUF quants range 1-bit (~270GB) to 8-bit (~870GB) — far beyond a single RTX 4090 at any quant level.
Quantizations
NVFP4 · GGUF
Fine-tune it
Permissive
QLoRA663.0 GB4× B200 192GB
LoRA2076.8 GBbeyond 8× B200
Full fine-tune15648.8 GBbeyond 8× B200
QLoRA SFT on ~10k samples ≈ $14.66 (4× B200 192GB)
Plan a fine-tune →
Inkling family
Elo progression across releases
InklingJul 20262194.1Inkling-SmallJul 20262366.6
Resources
HFthinkingmachines/InklingDOCShttps://thinkingmachines.ai/model-card/inkling/APIAvailable via Thinking Machines Lab API
API price $1.87/$4.68 · each benchmark row carries its own source badge (see methodology)