Thinking Machines Lab OPEN-WEIGHT · Apache 2.0 (open weights released on Hugging Face Nov 15, 2026; NVFP4 and MXFP8 checkpoints) · Context: Up to 1M tokens (per the open-weight release)

Thinking Machines Inkling 2 Small

Second-generation Inkling model: 594B-total / 17B-active MoE with vision, scoring 72.1% on DeepSWE v1.1, 86.3% on SWE-Bench Pro v2 (Hard) and 85.4% on MCP Atlas, with the top open-weight DesignArena website score; Apache 2.0 weights promised after red-teaming.

Provider
Thinking Machines Lab (Apache 2.0 (open weights released on Hugging Face Nov 15, 2026; NVFP4 and MXFP8 checkpoints))
Architecture
48-layer sparse MoE decoder (8 of 384 experts per token) with native text + image input (292M-parameter vision encoder) and variable effort; NVFP4 pre/mid/post-training
Parameters
594B total / 17B active
Context Window
Up to 1M tokens (per the open-weight release) (Max Output: 0 tokens)
API Pricing
Input: $0.00 / 1M · Cached: $0.00 / 1M · Output: $0.00 / 1M
VRAM (16K Ctx)
Q4_K_M: 359.03 GB · Q8_0: 632.32 GB · FP16: 1191.11 GB
Hardware Target
2× B300 or 4× H200 (NVFP4, ≥500 GB VRAM); MXFP8 needs ≥800 GB

Interactive Token Cost Estimator (Inkling 2 Small)

Est. API Cost (0% Cache): --/mo | With 70% Prompt Cache: --/mo
Compare all 12 models →

GPU Hardware Compatibility Matrix (Inkling 2 Small at 16K Context)

GPU / Hardware Profile VRAM Q4_K_M (4.8-bit) Q8_0 (8-bit) Est. Speed (Q4)
RTX 5060 Ti 16GB
448 GB/s
16 GB REQUIRES MULTI-GPU (359.03 GB) REQUIRES MULTI-GPU (632.32 GB) ~6 tok/s
RTX 4090 24GB
1008 GB/s
24 GB REQUIRES MULTI-GPU (359.03 GB) REQUIRES MULTI-GPU (632.32 GB) ~14 tok/s
RTX 5090 32GB
1792 GB/s
32 GB REQUIRES MULTI-GPU (359.03 GB) REQUIRES MULTI-GPU (632.32 GB) ~24 tok/s
M4 Max 64GB
546 GB/s
52 GB REQUIRES MULTI-GPU (359.03 GB) REQUIRES MULTI-GPU (632.32 GB) ~7 tok/s
2× 24GB (48GB)
920 GB/s
48 GB REQUIRES MULTI-GPU (359.03 GB) REQUIRES MULTI-GPU (632.32 GB) ~12 tok/s
H100 80GB
3350 GB/s
80 GB REQUIRES MULTI-GPU (359.03 GB) REQUIRES MULTI-GPU (632.32 GB) ~45 tok/s
M4 Ultra 192GB
819 GB/s
164 GB REQUIRES MULTI-GPU (359.03 GB) REQUIRES MULTI-GPU (632.32 GB) ~11 tok/s
8× H200 (1,128GB)
4800 GB/s
1128 GB FITS IN VRAM (359.03 GB) FITS IN VRAM (632.32 GB) ~307 tok/s

Architectural Strengths & Target Workloads

  • Long-horizon agentic coding and autoresearch (fine-tuned a 4B model autonomously for $426)
  • Strong vision and screen grounding (ScreenSpot Pro 83.8%)
  • Built to be fine-tuned on Tinker

Primary Engineering Workloads: Agentic Coding · Visual Web & Game Building · Custom Fine-Tuning Base.

Frequently Asked Questions (Inkling 2 Small)

How much does Thinking Machines Inkling 2 Small cost per 1M tokens?

Thinking Machines Inkling 2 Small is priced at $0.00 per 1M uncached input tokens, $0.00 per 1M cached input tokens, and $0.00 per 1M output tokens (Tinker (inference and fine-tuning); hosted APIs via NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, Cloudflare and others (per-provider pricing); self-host with SGLang, vLLM, TokenSpeed or Unsloth).

What is the context window of Thinking Machines Inkling 2 Small?

It supports up to Up to 1M tokens (per the open-weight release) (1,000,000 tokens) and a maximum generation output of 0 tokens.

Full TechPulse Engineering Analysis

Thinking Machines' Inkling 2 Small: a 594B MoE With 17B Active, Apache 2.0 Weights Promised

Mira Murati's lab ships its second model generation on Tinker first, with full weights due after outside red-teaming. Coding, vision and long-horizon autoresearch are the headline gains.

Read Full Technical Dispatch (5 min read) →
Ink-and-wash illustration of an ink bottle on a paper desk whose spilled ink forms a ladder climbing to an open window, with a small paper crane resting beside it