Model profile & tracker
Inkling 2 Mini: Price, Benchmarks, Local Setup
Inkling 2 Mini (released 20 November 2026): $0.15/$0.50 per million tokens, open weights, 1M context, 26.4B/3.4B MoE. Benchmarks, Ollama setup, vs Qwen3.8-27B.
Updated 22 November 2026. This page tracks Inkling 2 Mini — Thinking Machines Lab's open-weight model released on 20 November 2026 — and gets rewritten when the price changes, the Inkling 2 Small sibling ships, or the local-run requirements move.
Inkling 2 Mini is the smallest of the second Inkling generation: 26.4B total parameters, 3.4B active per token, Apache 2.0, a 1M-token context window, and a hosted price of $0.15 input / $0.50 output per million tokens. It is small enough to run on a Mac laptop and cheap enough to put behind a high-volume feature — the model you run on every request and keep the expensive one for the cases that fail.
The specs, as published
| Spec | Value |
|---|---|
| Model ID | thinkingmachines/inkling-2-mini |
| Released | 20 November 2026 |
| Licence | Apache 2.0, weights on Hugging Face and Ollama |
| Parameters | 26.4B total, 3.4B active (MoE) |
| Design | Looped transformer: 24 layers, the middle 12 run three times |
| Context window | 1,000,000 tokens |
| Max output | 1,000,000 tokens |
| Input price | $0.15 / 1M tokens |
| Cached input | $0.015 / 1M tokens |
| Output price | $0.50 / 1M tokens |
| Checkpoint size | About 20GB, NVFP4 with the last four layers in MXFP8 |
| Local hardware | Apple silicon with 32GB or more of memory, one NVIDIA DGX Spark |
| Input modalities | Text and image |
The looped design is the interesting part. Instead of a deeper network, the middle 12 of 24 layers run three times on each token: more computation per token, without more stored weights. That is how a 3.4B-active model keeps pace with dense models three times its size on the benchmarks below.
The benchmarks Thinking Machines published
| Benchmark | Inkling 2 Mini | Closest rival |
|---|---|---|
| SWEBench Verified (bash agent) | 79.2% | best in group |
| SWEBench Pro v2 (Hard) | 77.9% | Qwen3.8-27B: 73.8% |
| SWEBench Multilingual | 73.7% | top score listed |
| DeepSWE v1.1 | 48.2% | Qwen3.8-27B: 42.2% |
| Terminal-Bench v2.1 | 71.6% | Qwen3.8-27B: 73.0% |
The comparison group is Thinking Machines' own: Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL. These are the vendor's numbers on the vendor's chosen tests — the direction (a 3.4B-active model competing with 27–31B dense models) is well supported, the exact scores are not independently verified. Hands-on reports also note a gap between benchmark scores and live agentic performance: on one real debugging task the model looped for over an hour without finding the root cause.
Running it locally with Ollama
The one-line version, on a Mac with 32GB or more of memory:
ollama run inkling-2-mini
Ollama names the same minimum as Thinking Machines: Apple silicon with 32GB+. The 20GB checkpoint rules out smaller machines at full precision; Unsloth's compressed builds bring it to about 17.6GB at 4-bit and about 9GB at 1-bit, which puts 16GB and 24GB Macs in range with some quality loss. Beyond Ollama, the supported stack includes vLLM, SGLang, llama.cpp (GGUF), Hugging Face transformers, NVIDIA Dynamo and Exo Labs on Apple silicon.
Fine-tuning runs through Thinking Machines' Tinker platform, paid per token.
Inkling 2 Mini vs the other cheap options
vs DaisyRose 27B ($0.15 / $1.875): same input price, both Apache 2.0. Mini's $0.50 output is 3.75 times cheaper than DaisyRose's $1.875, and its 1M window is four times the 262K. DaisyRose remains the pick for multi-step reasoning; Mini is the pick for volume and very long inputs.
vs Qwen3.8-27B: the local model Thinking Machines measures against. Mini reads far fewer weights per token than a dense 27B, so local speeds should gap the same way. On the vendor's table Mini wins nearly every coding metric; Qwen3.8-27B keeps Terminal-Bench.
vs Gemini 4 Hydrogen ($1.50 / $7.50): not really a contest on price — Mini is a tenth the input cost and a fifteenth the output cost — but Hydrogen is a frontier-tier multimodal model and Mini is not. Compare them only if "cheap and open" is already settled.
FAQ
What does Inkling 2 Mini cost? $0.15 per million input tokens and $0.50 per million output tokens hosted, with cached input at $0.015. Self-hosting turns that into a GPU bill.
Can Inkling 2 Mini run on my Mac? On Apple silicon with 32GB of unified memory or more, yes, via Ollama. Below 32GB you need Unsloth's 4-bit (17.6GB, 24GB machines) or 1-bit (about 9GB, 16GB machines) builds, with quality loss.
Is Inkling 2 Mini open source? The weights are open under Apache 2.0. It is one of the two open-weight models in the comparison table, alongside DaisyRose 27B.
When was it released? 20 November 2026, alongside the Inkling 2 generation's larger siblings. The new models page tracks what shipped this month.
Should you use it? For classification, extraction, tagging, summarising and first drafts at volume: yes, it is the cheapest output in the table. For hard reasoning and code that has to pass tests, run it first and escalate the failures to a frontier model.