Inkling

Summary: Thinking Machines Lab's 975B total / 41B active parameter open-weight multimodal MoE model (Apache 2.0) with hybrid local/global attention, short convolutions, embedding RMSNorm, and learned relative-position bias instead of RoPE. 1M token context, supports BF16/MXFP8/NVFP4.

Details

Architecture

  • Type: Multimodal autoregressive transformer (decoder-only)
  • Layers: 66 decoder layers
  • Attention: Hybrid local/global
    • 55 local attention layers: 512-token sliding window
    • 11 global attention layers: full attention (learned bias applied to preceding 1024 tokens only)
  • MoE: Sparse Mixture-of-Experts
    • 256 experts, top-6 routing per token
    • 2 shared experts (always active)
    • 4.2% parameter activation per token (41B/975B)
  • Positional Encoding: Learned input-dependent relative-position bias (no RoPE)
  • Convolutions: Kernel-4 convolutions after K/V projections and on attention/MLP branch outputs
  • Normalization: Embedding RMSNorm (after token embeddings, separate from per-block pre-attention RMSNorm)

Parameters

  • Total: 975B
  • Active: 41B
  • Context window: 1M tokens
  • Modalities: Text, Image, Audio → Text output

Quantization & Deployment

  • BF16: 2 TB VRAM (8× B300 or 16× H200)
  • NVFP4 (W4A4): 600 GB VRAM (4× B300 SM100+, or 8× H200 W4A16)
  • Frameworks: SGLang, vLLM, TokenSpeed, Unsloth, Hugging Face Transformers
  • Weights: Hugging Face (thinkingmachines/inkling)
  • API: Tinker platform + third-party inference providers

Benchmarks (Model Card)

Category Benchmark Inkling Top Closed
Reasoning HLE (no tools) 29.7% GPT 5.6 Sol 47.2%
Reasoning HLE (with tools) 46.0% GPT 5.6 Sol 55.0%
Reasoning AIME 2026 97.1% GPT 5.6 Sol 99.9%
Reasoning GPQA Diamond 87.2% GPT 5.6 Sol 94.1%
Coding (agentic) SWE-Bench Verified 77.6% Claude Fable 5 95.0%
Coding (agentic) SWE-Bench Pro Public 54.3% Claude Fable 5 80.0%
Coding (agentic) Terminal Bench 2.1 63.8% GPT 5.6 Sol 89.5%
Agentic (general) GDPVal-AA v2 1238 Claude Fable 5 1760
Agentic (general) MCP Atlas 74.1% Claude Fable 5 83.3%
Factuality SimpleQA Verified 43.9% Gemini 3.1 Pro 77.3%
Vision MMMU Pro Standard 10 73.5% Gemini 3.1 Pro 82.0%
Vision CharXiv RQ (with Python) 82.0% GPT 5.6 Sol 87.8%
Audio VoiceBench 91.4% Gemini 3.1 Pro 94.3%
Safety FORTRESS Adversarial 78.0% Claude Fable 5 96.0%

Significance

Honest, broad benchmark profile — not benchmark-specialized. Interesting architectural departures from DeepSeek-V3/Kimi recipe: (1) short convolutions for local inductive bias, (2) embedding RMSNorm (unusual placement), (3) learned relative-position bias instead of RoPE (claim better extrapolation), (4) less extreme sparsity (4.2% vs 3.2%). Designed for fine-tuning via Tinker platform. Open weights (Apache 2.0) enables independent research.

Related Concepts

Sources

  • inkling-model-card: Model Properties — 975B/41B, 66-layer, hybrid local(55)/global(11) attention, 256 experts top-6 + 2 shared, 1M context
  • inkling-model-card: Numerics Support — BF16/MXFP8/NVFP4; NVFP4: 600GB VRAM (4×B300 or 8×H200)
  • inkling-model-card: Benchmarks — SWE-Verified 77.6%, Terminal Bench 63.8%, IFBench 79.8%, SimpleQA 43.9%, HLE no tools 29.7%
  • inkling-raschka-blog: Architecture observations — Kernel-4 convolutions, embedding RMSNorm, learned relative-position bias (no RoPE)
  • inkling-raschka-blog: Comparison — 4.2% active params vs Kimi K2.5's 3.2%; regular decoder vs Mamba-hybrid