HaiAI123

Curated Global AI Tools Directory

RisingIndustry update
↑ 19%

Thinking Machines Lab Open-Sources Inkling 2 Mini with 32GB Model Card Recommendation

11/21/2026 — 11/22, 09:13·10 sources (all)·18 reports (all)·In the 2026-11-22 briefing·In the 2026-11-21 briefing

Key takeaways

  • Thinking Machines Lab open-sourced the multimodal Inkling 2 Mini on November 20 under Apache 2.0, a looped Transformer with 26.4B total para…

Background & analysis

Thinking Machines Lab released the open-weights model Inkling 2 Mini on November 20, 2026, describing it as a fast, looped transformer small enough to run on a laptop, with weights of roughly 20GB and a separate variable-loop research variant. News spread on November 21 through a wave of account updates and media coverage. TestingCatalog reported 26.4B total parameters and 3.4B active parameters, with 5.4B weight activations per token to account for looping. Artificial Analysis framed the model as a 26B total, 5.4B active parameter mixture-of-experts model that scores 20 on the Intelligence Index.

On performance and cost, Artificial Analysis said Inkling 2 Mini matches Qwen3.8 27B (xhigh) and Nanbeige4.3-33B on the Intelligence Index with roughly 5x and 6x fewer active parameters. It reaches the same score with about 38% more output tokens and about 2.8x less compute, because 12 of its 24 layers run three times per token, and it averages about 152K output tokens per Intelligence Index task. A single task costs $0.59 and takes 9.0 minutes, versus $3.53 and 38.2 minutes for Qwen3.8 27B, with costs calculated from Thinking Machines' first-party pricing. The lab says it leads open-weights models under 100B parameters on DeepSWE v1.1 at 48.2%, another report puts it at 79.2% on SWE-bench Verified, and the model supports a 1M token context, text and image input, and 2.6 trillion training tokens.

Details diverge across coverage. TestingCatalog reported a roughly 20GB footprint and a 20.1 score on the AA Index. One DEV Community breakdown described 4.3-billion parameters with 14 attention layers repeated 22 times, while another matched the common account of 12 independent layers reused three times for 36 equivalent layers. Weights and inference code were published on Hugging Face under an Apache 2.0 license, with day-one support from SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo, and TokenSpeed, and inference plus fine-tuning on Tinker at $0.15 per million input tokens. The model card recommends 32GB or more of memory on Apple silicon, or 32GB or more of VRAM on other GPUs; the lab's demo ran offline on a 48GB MacBook Pro mid-flight, where researchers built a playable flight simulator in minutes.

AI-generated from 18 reports · updated 2 hours ago

Latest turnA hands-on analysis on Juejin describes Inkling 2 Mini, a 26.4B-parameter model that runs its middle layers three times, allowing it to fit into a laptop with 20GB of memory. The article frames this around local code models and the reasoning breakdown small models face in multi-file edits and agent tasks.

Related AI tools

updream
24-hour heatHeat index 596 · peak 681 · 10h ago
24 hours agonow

Reports on this story headlines open the original

Today
  1. A hands-on analysis on Juejin describes Inkling 2 Mini, a 26.4B-parameter model that runs its middle layers three times, allowing it to fit into a laptop with 20GB of memory. The article frames this around local code models and the reasoning breakdown small models face in multi-file edits and agent tasks.

    掘金AI score 58
Yesterday
  1. Thinking Machines Lab has released Inkling 2 Mini, an open-weights model that fits in about 20GB of memory and runs offline on a laptop; the lab says it leads every open-weights model under 100bn parameters on coding benchmarks such as DeepSWE. Its trick is looping — the middle layers run three times per token — inside a mixture-of-experts model with 26.4bn total and 3.4bn active parameters.

    The Next WebAI score 54
  2. Thinking Machines has released Inkling 2 Mini, a 26.4B-parameter looped MoE transformer with 3.4B active parameters, according to MarkTechPost. The model runs on a 32GB Mac or a single DGX Spark and scores 79.2% on SWE-bench Verified.

    MarkTechPostAI score 64
  3. Thinking Machines has released Inkling 2 Mini, an open looped transformer that runs 12 of its 24 layers three times per token, letting it reason internally for longer without emitting extra text. The Decoder reports it ties Qwen3.8 27B on Artificial Analysis's Intelligence Index at roughly one-sixth the cost.

    The DecoderAI score 59
  4. Thinking Machines Lab released Inkling 2 Mini on November 20 under Apache 2.0: a multimodal, looped MoE model with 26.4B total parameters, roughly 3.4B–5.4B active, and up to 1M context, pitched at laptops and local GPU rigs. This engineering breakdown covers the architecture, first-party API pricing, and early vLLM deployment notes, including kernel edge cases at launch.

    DEV Community · AIAI score 59
  5. Thinking Machines Lab released the open-weights model Inkling 2 Mini on November 20, saying in its blog post that it runs on a laptop with roughly 20GB of weights, with a demo shot offline on a 48GB MacBook Pro. The model card recommends 32GB or more of memory on Apple silicon and 32GB or more of VRAM on other GPUs.

    V2EXAI score 52
  6. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released Inkling 2 Mini, a 27B MoE model with 3.4B active parameters that fits in about 20GB of memory and runs locally on consumer laptops. The model uses a looping Transformer mechanism to think more and output less; researchers ran it offline on a MacBook Pro and built a playable flight simulator in minutes. It scores 48.2% on DeepSWE v1.1.

    智东西AI score 74
  7. Thinking Machines Lab has released Inkling 2 Mini, an open-weights model with 26.4 billion parameters under Apache 2.0 and a roughly 20GB footprint that runs offline on a laptop. It follows Inkling 2 Small, a 594-billion-parameter model released two weeks earlier that was impractical outside a data center. The author says they have not yet tested it themselves.

    DEV Community · AIAI score 68
  8. Thinking Machines Lab released Inkling 2 Mini, an open-weights multimodal model under Apache 2.0 with 26.4B total and 3.4B active parameters, using 12 unique layers looped three times for an effective depth of 36. It runs on 20 GB of memory with a context window of up to 1M tokens, and the lab reports 48.2% on DeepSWE v1.1, ahead of open models under 100B parameters.

    DEV Community · AIAI score 59
  9. Thinking Machines Lab released Inkling 2 Mini, an open-weights 4.3B-parameter looped-transformer model that runs on a laptop. Its 14 attention layers loop 22 times, and the Apache-2.0 model ships weights and inference code on Hugging Face with a 1M-token context; it costs $0.15 per million input tokens on Tinker.

    DEV Community · AIAI score 64
  10. Artificial Analysis says Thinking Machines has released Inkling 2 Mini, an open-weights looped transformer that scores 20 on its Intelligence Index. With 5.4B active parameters, it reportedly matches Qwen3.8 27B (xhigh) while using roughly 5x fewer active parameters.

    TechmemeAI score 53
  11. Thinking Machines Lab has released Inkling 2 Mini, a 26B open-weights looped transformer that runs on a laptop. It has 26.4B total parameters with 3.4B active, sits at roughly 20GB, and scores 20.1 on AA Index for agentic coding. Weights are on Hugging Face, with inference and fine-tuning live on Tinker and day-one support from SGLang, vLLM and Ollama.

    TestingCatalogAI score 65
  12. Thinking Machines Lab has released Inkling 2 Mini, an open-weights 26.4B-parameter transformer that loops its middle layers three times per token to get better reasoning with fewer tokens. The company says the model is fast and small enough to run on a laptop.

    TechmemeAI score 67
  13. Artificial Analysis reports that Inkling 2 Mini costs $0.59 and takes 9.0 minutes per Intelligence Index task, versus $3.53 and 38.2 minutes for Qwen3.8 27B, roughly 6x cheaper and 4x faster. Costs use Thinking Machines' first-party pricing, and GPT-6 Luna (max) scores 28 at $0.15 per task.

    X:Artificial Analysis (@artificialanlys)First-partyAI score 50
  14. At the same Intelligence Index score, Inkling 2 Mini uses about 38% more output tokens than Qwen3.8 27B but roughly 2.8x less compute, according to Artificial Analysis. Twelve of its 24 layers run three times per token, adding compute without adding output tokens, and it averages about 152K output tokens per task.

    X:Artificial Analysis (@artificialanlys)First-partyAI score 42
  15. Artificial Analysis says Inkling 2 Mini matches Qwen3.8 27B (xhigh) and Nanbeige4.3-33B on the Intelligence Index with roughly 5x and 6x fewer active parameters respectively. It sits on the Pareto frontier for intelligence versus both total and active parameters.

    X:Artificial Analysis (@artificialanlys)First-partyAI score 48
  16. Thinking Machines has released Inkling 2 Mini, an open-weights looped transformer mixture-of-experts model with 26B total and 5.4B active parameters. It scores 20 on the Artificial Analysis Intelligence Index and sits on the Pareto frontier for intelligence versus active parameters.

    X:Artificial Analysis (@artificialanlys)First-partyAI score 65
  17. TestingCatalog reports that Thinking Machines has announced Inkling 2 Mini, a new open-weight mixture-of-experts model with 26.4B total parameters, 3.4B active parameters, and 5.4B weight activations per token to account for looping. The account says the model leads across several comparisons.

    X:TestingCatalog (@testingcatalog)AI score 56

Other stories people are talking about

How is heat calculated?About the method

Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.

This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.

Surge
Discussion rising fast
New
First report within 6 hours
Rising
Still gathering discussion

Back to AI industry updates →