Thinking Machines open-sources Inkling 2 Mini, with day-one vLLM support
Understand this story
Thinking Machines released Inkling 2 Mini, an open-weight recurrent Transformer model under the Apache 2.0 license, supporting text and image input and a 1 million token context window, and small enough to run on a MacBook or DGX Spark. The Hugging Face blog describes it as a multimodal MoE model with about 26B physical parameters and 3B activated per token, offering expert weights in mixed NVFP4 and MXFP8 precision, and shipping with a DSpark draft model that trades a small amount of memory for speed. Artificial Analysis says it scores 20 on its intelligence index, placing it on the Pareto frontier of intelligence versus active parameters. vLLM announced same-day day-one support for the model and its variable-loop research checkpoint TML-d24-varloop-research on NVIDIA Blackwell, Hopper and DGX Spark, and provided the supporting image vllm/vllm-openai:inkling-2-mini. vLLM says loop_num_passes sets the number of loops (at least 1, defaulting to the checkpoint's 3), and DSpark must use the default value; both settings are rejected on fixed-loop checkpoints such as Inkling 2 Mini. The Hugging Face blog says transformers, llama.cpp and MLX support are also provided, and SGLang offers day-one inference support on data center hardware. Under Thinking Machines' first-party pricing, each 1 million input, output and cache-hit tokens cost $0.15, $0.50 and $0.015 respectively.
AI overview based on source reports · Updated 56 min ago
Developments
2 developments
- Nov 20 08:00 · 1 reportsvLLM supports Inkling 2 Mini inference on day onevLLM Blog:Inkling 2 Mini on vLLM: Day-0 Support for a Looped Model
- Nov 20 08:00 · 2 reportsThinking Machines releases open-weight model Inkling 2 MiniArtificial Analysis Articles:Thinking Machines has released Inkling 2 Mini, an open weights looped transformer
Coverage timeline
Follow the coverage to explore different perspectives.
- Artificial Analysis ArticlesFeaturedThinking Machines has released Inkling 2 Mini, an open weights looped transformer
Thinking Machines released Inkling 2 Mini, an Apache 2.0 open-weights mixture-of-experts model with 26B total and 5.4B active parameters (including three passes through its looped layers, 3.1B unique weights per text token), scoring 20 on the Artificial Analysis Intelligence Index. It matches Qwen3.8 27B (xhigh) and Nanbeige4.3-33B on the index with roughly 5x and 6x fewer active parameters, supports text and image input with a 1M token context window, and is small enough to run on a MacBook or DGX Spark. It uses about 2.8x less compute and is about 6x cheaper and 4x faster per task than Qwen3.8 27B ($0.59 and 9.0 minutes vs. $3.53 and 38.2 minutes).
- vLLM BlogInkling 2 Mini on vLLM: Day-0 Support for a Looped Model
Thinking Machines announced Inkling 2 Mini, a multimodal MoE with 26B total and 3.5B active parameters, up to 1M tokens of context and a native DSpark speculator, whose middle 12 layers run three passes with shared weights, each pass keeping its own KV cache and reading gated messages from the pass before. vLLM supports it and the variable-loop research checkpoint TML-d24-varloop-research on day 0 on NVIDIA Blackwell, Hopper and DGX Spark, registering each extra pass as its own attention layer and fusing message reads, controllers and sublayer kernels. A Docker image, vllm/vllm-openai:inkling-2-mini, is provided, and vLLM serves the model for reinforcement learning with SkyRL.
- Hugging Face BlogFeaturedWelcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup
Thinking Machines Lab's Inkling 2 Mini is a multimodal Mixture-of-Experts model with about 26B total and 3B active parameters per token, using 24 decoder layers where the 12 middle layers run three times in a loop, plus a 1M-token context and image and text inputs. It ships in mixed precision with NVFP4 and MXFP8 expert weights and an optional DSpark drafter, with support in transformers, llama.cpp/Llama App, MLX and SGLang, and a variable-loop research variant is also released. The post reports it leads in 32 of 48 benchmarks against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL, and shares vibe tests including shader scenes, a Van Gogh recreation and a napkin-sketch-to-app build.
Activity for this story
Not enough continuous observations to show a trend yet.