All posts
Open Source Published Updated 11 min

November 20 agent guide: Inkling 2 Mini open weights

November 20, 2026 extended Thinking Machines’ Inkling 2 line downward. Inkling 2 Mini ships Apache 2.0 open weights as a 26.4B MoE looped transformer aimed at laptop-scale agentic coding with strong DeepSWE and Vals scores among sub-100B open models.

Part ofAgentic AI
Jigar JoshiJigar JoshiAgentic AI Architect and Consultant
November 20, 2026 cover for Inkling 2 Mini open weights
In this post (6 sections)

November 20, 2026 is an open-weight local-agent day. Thinking Machines Lab released Inkling 2 Mini eight days after Inkling 2 Small’s Apache 2.0 drop, targeting laptops and small GPUs without giving up agentic coding strength.

Architecture

Inkling 2 Mini stores 24 layers and runs the two middle blocks three times per token, so 24 stored layers behave like 48 depth for reasoning. Specs: 26.4B total MoE parameters, 3.4B active, 5.4B weight activations per token with looping, multimodal image-text-to-text tagging on Hugging Face, and native NVFP4 training/inference. Quantized GGUF paths support Ollama and llama.cpp.

Monitorability angle

The lab stresses that looping can improve answers without writing long chain-of-thought, which changes how teams oversee agents. Factual benchmarks do not improve with more loops the way long-context graph tasks can. Mini leans on search and tools (BrowseComp, DeepSearchQA, MCP Atlas) to cover knowledge gaps.

Who should adopt it

  • Teams that need offline or air-gapped coding agents after November 12’s Small weights proved the Inkling 2 recipe.
  • Platforms standardizing on Apache 2.0 MoE workers beside Qwen3.8-27B and Gemma 4 mid-size opens.
  • Eval groups measuring tokens per resolved DeepSWE or Terminal-Bench task on a single workstation.

Builder checklist

  1. 01
    Pull thinkingmachines/Inkling-2-Mini
    Confirm Apache 2.0 terms and NVFP4 vs GGUF path for your hardware.
  2. 02
    Benchmark against Qwen3.8-27B on the same harness
    Use DeepSWE and a private merge suite; do not trust only published scatter plots.
  3. 03
    Wire MCP tools early
    Mini’s knowledge gap is intentional; tool use is part of the product story.
  4. 04
    Keep Small for heavier hosts
    Mini does not replace Inkling 2 Small’s larger deployment profile.

Conclusion

November 20 gives builders a laptop-scale Inkling 2 worker with Apache 2.0 weights and a looping design aimed at agentic coding efficiency. Bacancy Technology’s rule: treat Mini as a local open lane beside Qwen4-Max API and GPT-6.2 Sol cloud defaults.

Need local open-weight agent hosting?

Bacancy Technology designs open-weight serving, NVFP4 capacity plans, and MCP tool allowlists. Consulting: /consulting/ . Contact: /contact/ .

Sources: thinkingmachines.ai — inkling 2 mini ; Hugging Face .

The weekly take

Agentic AI patterns, delivered Thursdays

What I am shipping, watching, and pruning out of client stacks each week. One email. No fluff.

Shipping an agentic AI project this quarter?
Book a 30-min consult
Frequently asked

Questions readers ask about this post

Share this post
LinkedIn Facebook WhatsApp