Skip to content
NewA calmer, faster ZeroHour
StoryModel releaseinfo

Thinking Machines releases Inkling 2 Mini; GGUF builds follow

4 sources · 6 articlesFirst reported Updated ()1
What's new

November 21: The Decoder added DeepSWE, Vals Index, and Artificial Analysis scores. AtomicChat’s November 20 uncensored LoRA and GGUFs report zero refusals on its own tests.

Merged summary · x-ai/grok-4.7 · rewritten as coverage arrivesSummaries by Vani AI

Inkling 2 Mini is out under Apache 2.0 with official weights, Unsloth GGUFs, Tinker access, and an AtomicChat uncensored build.

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, an Apache 2.0 model that accepts text and images and generates text. It is a 24-layer decoder-only transformer with sparse mixture-of-experts and looped middle layers, with Hugging Face weights, Tinker access, and local inference on listed NVIDIA and Apple hardware. Size listings conflict: one Hugging Face card says 26.4 billion parameters and Unsloth’s says 26.2 billion, both with 3.4 billion active, while The Decoder reports 26.4 billion total and 5.4 billion active per token. The Decoder says the company claims it leads open models under 100 billion parameters on agentic coding tests. Unsloth published GGUFs, and AtomicChat released an uncensored LoRA and GGUFs that, on its own tests, cut harmful-prompt refusals to zero while missing MMLU and KL gates.

Key facts
  • Each token routes to 8 of 128 experts; context up to 1 million tokens.
  • Three loop passes yield about 5.4 billion effective active parameters.
  • The Decoder: 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2.
  • AtomicChat: English refusals fell from 81 of 81 to 0; MMLU from 80.65% to 79.15%.
  • Uncensored GGUFs need llama.cpp b12600 or newer and merged PR #32694.
  • Unsloth UD-Q4_K_XL is 17.63 GB and 98.0% same-top-token versus Q8.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

Timeline

dated events in the reporting · oldest first
  1. Nov 20, 2026Thinking Machines released Inkling 2 Mini under Apache 2.0. · Hugging Face trending models
  2. Nov 20, 2026Unsloth published GGUF quantizations of Inkling 2 Mini. · Hugging Face trending models
  3. Nov 20, 2026AtomicChat released an uncensored LoRA and three GGUFs. · Hugging Face trending models
  4. Nov 21, 2026The Decoder reported coding scores and a 1-million-token window. · The Decoder

What each outlet adds

beyond the shared facts
  • Hugging Face trending modelsOfficial card: 26.4B, loops, experts, hardware, and Apache 2.0.
  • Unsloth GGUF listingLists 26.2B total and a 17.63 GB UD-Q4_K_XL file.
  • AtomicChatUncensored LoRA and GGUF refusal, MMLU, and llama.cpp requirements.
  • The DecoderSays 12 of 24 layers rerun and cites coding and AA scores.

Coverage

oldest first · one row per article · UTC
  1. · 5d ago Hugging Face trending modelsFirst report

    Thinking Machines Lab releases open-weight Inkling 2 Mini

    The Apache 2.0 checkpoint is about 20GB, with 26.4 billion parameters and 3.4 billion active per token.

  2. · 2d ago Hugging Face Blog

    Inkling 2 Mini announced for local-setup performance

    Only the Hugging Face post title is available, with no specs, license, or release status.

  3. · 2d ago Hugging Face trending models

    Thinking Machines Lab releases Inkling 2 Mini open weights

    The 26.2 billion-parameter multimodal model has 3.4 billion active parameters and an Apache 2.0 license.

  4. · 1d ago Hugging Face trending models

    AtomicChat posts uncensored Inkling-2-Mini GGUF builds

    AtomicChat says its adapter drops refusal to zero on its harmful tests while missing its MMLU and first-token KL gates.

  5. · 1d ago The Decoder

    Thinking Machines releases open Inkling 2 Mini looped model

    The Apache 2.0 model has 26.4 billion parameters, 5.4 billion active per token, and a 1-million-token window.

  6. · 1d ago MarkTechPost

    Thinking Machines introduces Inkling 2 Mini, a 26B looped MoE

    The captured page is a cookie notice, so only the title's claim that it runs on a 32GB Mac is available.