
Thinking Machines Lab released Inkling 2 Mini, an open-weights model that takes text and images, responds in text and fits in a laptop’s memory: the weights take up about 20GB. It is built for agentic work – coding, tool-assisted search and document tasks, including on your own machine. On SWE-bench Verified, according to the developer, Inkling 2 Mini solves 79.2% of tasks, against 77.8% for Qwen3.8-27B and 46.1% for Gemma 4 31B. On DeepSWE v1.1 the gap is wider: 48.2% versus 42.2% for Qwen3.8-27B.
According to Artificial Analysis data that OpenRouter lists, Inkling 2 Mini scores 20.1 on the Intelligence Index – the same as Qwen3.8-27B. But Qwen3.8-27B is a dense model that uses all 27B parameters for every token, while Inkling 2 Mini uses 5.4B, counting the repeated passes. That comes down to a mixture of experts and a loop of layers.
What’s new in Inkling 2 Mini
The main novelty is the loop. A standard transformer passes each token through a stack of layers once; Inkling 2 Mini runs its middle layers three times. The model does more “thinking” internally without writing its reasoning out as text, and, the developer says, reaches an answer in fewer tokens – especially in agentic coding.
The second is the number format. Most of the weights were trained and released in 4-bit NVFP4, and the last four layers in 8-bit MXFP8. The model was not compressed after training, so, as the developer stresses, there is no conversion loss from quantization.
Third, the company released a research variant, TML-d24-varloop-research, trained to run with anywhere from one to five passes. It was used to probe the risk that comes with looping: the model can reason “silently”, and an observer reading its chain of thought would not see it. With five passes, reasoning was less than half as long as with one – 1,661 tokens versus 3,703 on average – yet stayed just as legible.
How the model is built
The model has 24 layers in four blocks of six. The first and last blocks run once, and the two middle blocks run three times with shared weights, giving 48 layer applications per token. According to the model card, the layers are a mixture of 128 experts, each token is routed to eight of them, and the context window is up to 1M tokens. In each block, as Hugging Face describes, five layers see a window of the last 1,024 tokens and one sees the full context.
The model started with 384 experts. In midtraining the company pruned two-thirds of them, keeping the 128 most-used experts in each layer, and recovered the lost quality by training the model on 1T tokens distilled from a larger internal Inkling model. Reinforcement learning took another 2.2T tokens. By the company’s numbers, the loop pays for itself: in an experiment with unpruned versions, the non-looping model cost 7.5% less compute to train but needed 140% more inference compute to reach the same loss on 512K-token agentic traces.
Inkling 2 Mini vs other open models
Thinking Machines compares the model with other open models: the similarly sized dense Qwen3.8-27B and Gemma 4 31B, and Nemotron 3.5 Super VL, which has almost five times as many parameters. The independent Artificial Analysis index shows a tie with Qwen3.8-27B, though Qwen reaches its 20.1 only at Xhigh reasoning effort; at Medium it scores 17.5.
| Metric | Inkling 2 Mini | Qwen3.8-27B | Nemotron 3.5 Super VL | Gemma 4 31B |
|---|---|---|---|---|
| Parameters: active / total, B | 5.4 / 26.4 | 27 / 27 | 12 / 124.5 | 30.7 / 30.7 |
| Artificial Analysis Index v5.0, points | 20.1 | 20.1 | 13.6 | 4.4 |
| SWE-bench Verified, % | 79.2 | 77.8 | 71.1 | 46.1 |
| DeepSWE v1.1, % | 48.2 | 42.2 | 24.7 | 0.3 |
| Terminal Bench v2.1, % | 71.6 | 73.0 | 60.7 | 43.4 |
| BrowseComp, % | 79.0 | 61.6 | 78.9 | 37.0 |
| AutomationBench-AA, % | 41.8 | 48.2 | 14.8 | 4.9 |
| AA-Omniscience: non-hallucination rate, % | 63.4 | 71.4 | 69.5 | 14.5 |
Data from Thinking Machines. Active parameters for Inkling 2 Mini include the three passes; without them, the figure is 3.4B.
Where the model shines and where it falls short
Inkling 2 Mini’s strengths are long-running agentic work, search and long documents. On the BrowseComp agentic search benchmark it scores 79.0% versus 61.6% for Qwen3.8-27B, and on the long-context LongBench v2 – 66.9% versus 52.3%. It is also efficient: Stanford researchers, using their methodology, rated its “intelligence per joule” on DGX Spark at 0.703 versus 0.430 for Qwen3.6-35B-A3B, although Inkling ran in MXFP8 with low reasoning effort and Qwen in NVFP4. That is why the developer pitches it for running several agents at once on a single machine.
It loses on office work and knowledge. In all five Artificial Analysis domain indexes where both models appear, Qwen3.8-27B at Xhigh is slightly ahead, most visibly in strategy and operations: 39.0 versus 35.2. Looping does not add room for facts – knowledge is mostly a function of total parameter count, the developer explains – and Inkling 2 Mini’s non-hallucination rate on AA-Omniscience is lower: 63.4% versus 71.4%. The company itself admits that stronger small models exist for multi-turn chat and instructions with many constraints: Inkling 2 Mini scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1.
Pricing, access and running it locally
Inkling 2 Mini’s weights are open under the Apache 2.0 license. Via API, the model is available on Tinker, Thinking Machines’ platform for inference and fine-tuning, and on OpenRouter at $0.10 per 1M input tokens, $0.40 per 1M output tokens and $0.02 for cache reads. There it is served by a single provider so far, Prime Intellect, at 73 tokens per second, while Qwen3.8-27B has 24 providers, the fastest of which deliver 303–357. Qwen3.8-27B is listed there at $0.035 for input, $1.50 for output and $0.0175 for cache reads, though prices vary across providers.
Our estimate: an agent session with 2M input tokens, 90% of them read from cache, plus 100K output tokens would cost about $0.10 with Inkling 2 Mini (0.2M × $0.10 + 1.8M × $0.02 + 0.1M × $0.40) and $0.19 with Qwen3.8-27B (0.2M × $0.035 + 1.8M × $0.0175 + 0.1M × $1.50) – almost twice as much. Most of the savings come from the cheap output.
Locally, the model runs via Ollama on any Mac with at least 32GB of memory and on NVIDIA DGX Spark; vLLM, SGLang, llama.cpp and MLX support it as well. Unsloth has published smaller quantized versions: the 4-bit UD-Q4_K_M weighs 16.61GB and picks the same top next token as the 22.87GB 8-bit version 97.9% of the time. The ggml-org build launches with a single command: llama serve -hf ggml-org/Inkling-2-Mini-GGUF:IQ4_NL.
The bottom line
Inkling 2 Mini suits anyone who wants an open agent on their own computer: it is on par with Qwen3.8-27B at coding, noticeably stronger at search and long documents, and runs on a Mac with 32GB of memory. Via API, by our estimate, an agent session costs almost half as much as with Qwen3.8-27B, but the model has only one provider so far, and it is much slower than Qwen’s fastest providers. For office work and fact-based questions, Qwen3.8-27B is the better fit: it hallucinates less and, at Xhigh, is slightly ahead on the domain indexes too. In multi-turn chat and instructions with many constraints, Inkling 2 Mini, as the developer itself admits, also trails some small models.
