Thinking Machines Lab releases Inkling 2 Mini: 26B-parameter looped MoE with 3B active and 1M-token context
Thinking Machines Lab has released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B physical total parameters and 3B active per token. It repeats its middle layers three times, supports a 1M-token context, and ships with a DSpark drafter. Hugging Face reports it leads 32 of 48 benchmarks against four comparable models.
Thinking Machines Lab has released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B physical total parameters and 3B active parameters per token, designed for local setups. It supports a 1M-token context and accepts text and image inputs. Hugging Face announced the release on November 20, 2026, in a post co-written by Merve, Pedro Cuenca, Aritra Roy Gosthipaty and Guarin Flück.
Architecture: a looped transformer
Inkling 2 Mini has 24 physical decoder layers. Layers 0-5 run once, layers 6-17 run three times, and layers 18-23 run once. That gives 48 logical layers (6 + 3x12 + 6) from 24 physical ones. All three passes through layers 6-17 reuse the same weights.
Key specifications:
- Hidden size: 2,304 (Inkling 2 Small uses 4,608)
- Experts: 8 routed per token from 128 SwiGLU experts per layer (Small selects from 384)
- Expert intermediate width: 1,152
- Attention: 24 query heads, 8 KV heads; each six-layer block has five sliding-window layers (window of up to 1,024 tokens) and one full-attention layer
- Layer visits: 48 per token, the same as Inkling 2 Small
Passes communicate through cross-pass messages. Each looped layer reads a message from the previous pass's outputs using its own learned query. A small gating controller then decides how much of the current input to forget and how much of the message to add. Reusing core weights reduces the number of distinct layers stored in the checkpoint, but all 48 visits still execute and each keeps its own KV and short-convolution state.
Weights, drafter and research variant
Thinking Machines Lab ships the model in a mixed-precision format with NVFP4 and MXFP8 expert weights. A DSpark drafter is included for additional speed at a small memory cost. A research checkpoint, TML-d24-varloop-research, shares the architecture but was trained to accept a variable number of loop passes. According to Hugging Face, setting 1 or 2 passes instead of 3 speeds up inference with slightly lower quality. The team says it tested the variant but has not evaluated it in depth.
The model is available as thinkingmachines/Inkling-2-Mini. Reasoning effort can be set from 0.0 to 0.99 through the chat template. The named levels are none (0.0), minimal (0.1), low (0.2), medium (0.7), high (0.9, default), xhigh (0.99) and max (0.99). API pricing is not yet disclosed.
Benchmarks
Hugging Face states that Inkling 2 Mini is state-of-the-art in its class on the Artificial Analysis Intelligence Index v5. Against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL, it leads on 32 of 48 benchmarks. Categories include coding and autoresearch, agentic search and tool use, vision, reasoning and knowledge, and safety and alignment.
On coding, it leads FrontierCode v2.0, ties Qwen3.8-27B on Terminal-Bench 4.0, and trails Qwen on AutomationBench-AA v1.0.6. Full scores are on the Inkling 2 Mini model card. The post does not list individual numeric scores in the text covered here. In Hugging Face's own informal tests, the model is at least comparable with Qwen 3.8 27B, though the authors say more validation is needed.
Inference support
Support is available in transformers, Llama App (llama.cpp) and MLX. SGLang provides day-zero inference support for production deployments on datacenter hardware.
What this means
Inkling 2 Mini gets more compute per token by reusing weights across passes, instead of storing more distinct layers. With 3B active parameters and a 1M-token context, it targets local hardware. The mixed NVFP4/MXFP8 expert weights and the bundled drafter are aimed at the same goal. The variable-loop research checkpoint also offers a way to trade quality for speed at inference time, though it has not been evaluated in depth. The benchmark claims come from Hugging Face's summary and should be checked against the model card and independent runs.
Related Articles
Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context
Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.
Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license
Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.
Inworld releases Voice Isolator v1: Apache-2.0 48 kHz voice isolation models at 64M and 58M parameters
Inworld has released Voice Isolator v1, two Apache-2.0 models for 48 kHz voice isolation: a 64M-parameter offline model and a 58M-parameter causal realtime model. The realtime model has 63 ms algorithmic latency at 20 ms of look-ahead. Both do denoising only; restoration is planned for version 2.
Cloudflare releases open-weight Clairic 27B and new Clairic-flash 9B for data extraction
Cloudflare announced on November 18, 2026 an updated 27-billion-parameter Clairic and a new 9-billion-parameter Clairic-flash. Both have open weights on Hugging Face and are available as hosted Workers AI endpoints.
Comments
Loading...