Inkling 2 Small review: Thinking Machines Lab model features, pricing and open weights plans

Inkling 2 Small from Thinking Machines Lab delivers a 594B-parameter sparse MoE model with 17B active parameters per token and multimodal input on the Tinker platform. Launched on November 5, 2026, it scores 31 on the Artificial Analysis Intelligence Index v5.0.1, an 18-point gain over the prior Inkling-Small generation. The model supports up to 1M-token context, PNG and JPEG image input, and text output only.

Thinking Machines Lab positions Inkling 2 Small for agentic coding, vision-driven design tasks, and enterprise knowledge work. It currently runs exclusively on Tinker with serverless inference in beta. Full weights under Apache 2.0 are scheduled after external safety testing, expected around mid-November 2026.

Inkling 2 Small features and performance benchmarks

The architecture uses 48 layers with hybrid local and global attention plus a 292M-parameter vision encoder that processes 20×20 patches. Input images should fall between 40 and 4096 pixels. Variable reasoning effort lets users trade speed for depth on complex queries. The 1M-token context window supports long documents or multi-turn agent sessions, though Tinker currently offers 64K and 256K variants for most users.

Benchmark results show clear gains in vision and agentic tasks. PerceptionBench reaches 64.7 percent, MMMU Pro hits 81.6 percent, and ScreenSpot Pro demonstrates strong screen understanding. On enterprise suites, EnterpriseBench-CoreCraft scores 71.8 percent and JobBench reaches 70.5 percent. Coding performance stands at 86.3 percent on SWE-Bench Pro v2 Hard, while terminal and command-line benchmarks remain weaker than leading closed models.

• Input: text plus images (PNG/JPEG)
• Output: text only
• Context: up to 1M tokens
• Parameters: 594B total, 17B active
• Vision encoder: 292M parameters

Compared with its predecessor, the model reduces hallucination rates through unified reinforcement learning after efficient pre-training. Users report more reliable outputs on iterative design and knowledge retrieval workflows.

Thinking Machines Inkling pricing and availability

Tinker charges $0.30 per million input tokens and $1.00 per million output tokens, with cached input at $0.03. Training runs cost $1.50 per million tokens at 64K context or $3.00 at 256K context. Storage adds $0.10 per GB per month. Serverless inference remains in beta and is available immediately for the model.

The service includes fine-tuning tools and a Playground interface. No self-hosted weights or third-party API endpoints exist yet. Access is limited to the Tinker platform for both inference and training workloads.

These rates position the model competitively for high-volume vision and agent workloads, though final open-weights pricing from other hosts remains unknown.

Open weights model plans and next steps

Thinking Machines Lab follows a staged safety process before releasing full weights under Apache 2.0. External red-teaming must complete before the planned mid-November 2026 drop. Fine-tuned checkpoints will accompany the base weights once released.

Broader platform availability is expected in the weeks after the weights release. Independent evaluations will likely follow, particularly on coding and terminal benchmarks where current scores lag. Organizations planning deployments should monitor the official announcement for exact timing and any usage restrictions tied to the open license.