Unofficial guide
Inkling 2 Mini vs Inkling 2 Small
Last updated
Inkling 2 Mini is the small local model. Inkling 2 Small is the 594 billion parameter model released on 5 November 2026. Mini, released on 20 November 2026, keeps the same broad design (text and images in, a 1 million token context, 8 experts per token, Apache 2.0) and cuts the stored model down to 26.4 billion parameters by pruning experts and looping its middle layers.
This comparison is unofficial. It uses the two model cards and the two config.json files on Hugging Face. It does not invent a head-to-head score where Thinking Machines did not publish Small's number next to Mini's.
The differences that change a buying decision
| Inkling 2 Mini | Inkling 2 Small | |
|---|---|---|
| Released | 20 November 2026 | 5 November 2026 |
| Total parameters | 26.4 billion | 594 billion |
| Active parameters | 3.4 billion, 5.4 billion effective with 3 passes | 17 billion |
| Stored layers | 24, hidden size 2304 | 48, hidden size 4608 |
| Layer visits per token | 48 (middle 12 layers run 3 times) | 48 (one visit each, no loop_config) |
| Experts | 8 of 128, no shared experts | 8 of 384, no shared experts |
| Expert width | 1152 | 2304 |
| Attention | 24 query heads, 8 KV heads | 48 query heads, 8 KV heads |
| Sliding window | 1024, plus one full layer per block of six | 1024, same five-local-to-one-full pattern |
| Context | 1,048,576 tokens | 1,048,576 tokens |
| Inputs | Text and images, text out | Text and images, text out |
| Vision tower | 23 layers, hidden size 1024, 16 heads | 23 layers, hidden size 1024, 16 heads |
| Numerics on the release checkpoint | NVFP4, last four layers MXFP8. About 20 GB. | MXFP8 and NVFP4. The card does not state a download size in GB. |
| Where it is meant to run | 32 GB+ Apple silicon, DGX Spark, or 32 GB+ VRAM recommended | Not described as a laptop model |
| License | Apache 2.0, plus the lab's acceptable use policy | Apache 2.0 |
Mini started training wider than it shipped. The announcement says it began at sparsity 48, choosing 8 experts out of 384, which is Small's expert count, and that mid-training pruned each layer down to the 128 experts that were used most. Healing was 1 trillion tokens distilled from a larger internal Inkling model, then SFT on a mix similar to Small's, then reinforcement learning of 2.2 trillion tokens over 2,000 steps, plus the same style of on-policy distillation used for Small.
What Thinking Machines says about quality
The Mini model card is direct: Mini is less capable than Small across most of their evaluations, and the largest gaps are factual recall and knowledge-heavy questions. Looping does not add facts. The announcement says extra passes helped long-context tasks such as GraphWalks and MRCR v2, and did not help AA-Omniscience, SimpleQA Verified, SimpleVQA, or WorldVQA.
They do claim Mini gets close on one coding-systems benchmark. On PMPP-Hard (AUC over 0–10 hours, pi v0.83.0) Mini scores 455.82, and the announcement says that approaches the much larger Small. Small's own PMPP number is not printed in that sentence, so it is not printed here.
On chain-of-thought control, the announcement compares them in one band. Among chains of 750–1,250 tokens, Mini and the variable-loop research model at 3 passes follow a "control the chain of thought" instruction at most 0.1% of the time, against 0.3% for Inkling 2 Small and 1.3% for Qwen3.8-27B. That is a monitorability measurement, not a coding score. Lower means the model is worse at steering its written reasoning to satisfy a side instruction.
Other models, only where a source put them in the same table
The Mini model card's comparison set for the main score table is Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL. A wider cut of the card adds further columns, including Beam 1.0. The rows below are the ones people actually ask about when they are choosing a local coding model. Scores are the lab's. The announcement says evaluation setups vary.
| Inkling 2 Mini | Qwen3.8-27B | Gemma 4 31B | Nemotron 3.5 Super VL | |
|---|---|---|---|---|
| Active / total (B) | 5.4 / 26.4 | 27 / 27 | 30.7 / 30.7 | 12 / 124.5 |
| DeepSWE v1.1 | 48.2% | 42.2% | 0.3% | 24.7% |
| SWE-bench Verified, bash only | 79.2% | 77.8% | 46.1% | 71.1% |
| SWE-bench Pro v2 Hard | 77.9% | 73.8% | 14.2% | 44.1% |
| Terminal-Bench v4.0 | 5.6% | 5.6% | 0.0% | 0.0% |
| Terminal-Bench v2.1 | 71.6% | 73.0% | 43.4% | 60.7% |
| AA Index v5.0 | 20.1 | 20.1 | 4.4 | 13.6 |
| AA-Omniscience score | -8.6 | -6.4 | -48.5 | -2.9 |
| MMMU Pro | 74.9% | 76.3% | 73.4% | 72.3% |
Muse Glimmer is in the same source table at 30B active and total. On SWE-bench Verified it is 76.0%, on DeepSWE 4.4%, on Terminal-Bench v2.1 51.7%. It is omitted above only so the table stays readable. The model card has every cell.
Read the Terminal-Bench pair as two different tests. v4.0 with mini-swe-agent is 5.6% for both Mini and Qwen, a low absolute number. v2.1, with Mini running in pi v0.83.0, is 71.6%. Quoting one of those without the version is how comparison posts go wrong.
The Hugging Face post's vibe tests say Mini felt at least comparable to Qwen 3.8 27B and that more validation is needed. That is their impression, not an extra benchmark. Stanford's intelligence-per-joule figures in the announcement use yet another set of models (Qwen3.6-35B-A3B, Gemma-4-26B-A4B, Qwen3.5-27B) at mixed precisions. Those numbers are on the overview, with the precision caveat.
Picking one
- You have a 32 GB Mac or you are trying to stay near 20–24 GB of memory: Mini, official checkpoint or a 4-bit GGUF. Small does not have a laptop story in its model card.
- The job is closed-book facts, long chat, or instructions with many constraints: the Mini card says this is where a small model gives ground. Qwen3.8-27B ties or leads several of those rows in the lab's own table.
- The job is agentic coding with search and tools, on one machine: Mini is the model the announcement is arguing for, and the DeepSWE and SWE-bench Pro rows are the evidence they lead with.
- You need to study loop depth: download
TML-d24-varloop-research, not Small and not the fixed Mini checkpoint.
Memory for Mini is spelled out on requirements. Nothing on this page is a reason to expect Small's weights to load on a laptop.
Common questions
Is Inkling 2 Mini the local version of Inkling 2 Small?
That is how Thinking Machines and Hugging Face describe the relationship. Mini is the checkpoint with a ~20 GB footprint aimed at laptops, Apple silicon with 32 GB or more, and a DGX Spark. Small is a 594 billion parameter model. The Mini model card says Mini is less capable than Small across most of their evaluations, especially factual recall.
Is Inkling 2 Small a looped transformer?
Small's model card calls it a 48-layer decoder-only mixture-of-experts and does not call it looped. Its config.json has no loop_config. Mini's config is model_type inkling2_loop, with layers 6–17 run three times. Both designs apply about 48 layer visits per token. Small stores 48 layers. Mini stores 24 and reuses the middle 12.
Does Inkling 2 Mini beat Qwen3.8-27B?
On some rows Thinking Machines published, yes. On others, no. DeepSWE v1.1 is 48.2% versus 42.2%, and SWE-bench Pro v2 Hard is 77.9% versus 73.8%. Terminal-Bench v2.1 is 71.6% versus 73.0%, and AA-Omniscience is -8.6 versus -6.4. Harnesses differ by row. This site did not re-run them.
Sources
- Thinking Machines Lab, “Introducing Inkling 2 Mini” Announcement dated 20 November 2026.
- Inkling 2 Mini model card Hardware, license, architecture, evaluations, and limitations.
- Inkling 2 Small model card 594B total, 17B active, 48 layers, 8 of 384 experts. Released 5 November 2026.
- Thinking Machines Lab, “Introducing Inkling 2 Small”
- Inkling 2 Mini config.json Hidden size, layer count, loop bounds, experts, and context length.
- Inkling 2 Small config.json
- Hugging Face blog, “Welcome Inkling 2 Mini” Published 20 November 2026. Transformers, llama serve, SGLang, and MLX commands.
- Hugging Face model API for Inkling-2-Mini Reports 26,388,646,440 parameters, in line with the 26.4B figure.