Tasks and inputs
Inkling 2 Mini is listed for Image-text-to-text, Text generation, Coding & Code Generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 21, 2026
Inkling 2 Mini is Thinking Machines' Apache 2.0 looped MoE model (26.4B total, 3.4B active) for agentic coding that runs on a 32 GB laptop.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights (Apache 2.0). Hosted on Tinker; publisher lists no per-token price on the announcement.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Can a model that fits in about 20 GB do serious agentic coding? Inkling 2 Mini, released by Thinking Machines Lab on 20 November 2026, is a direct attempt at a yes. It is a 26.4B-parameter mixture-of-experts model with 3.4B active, it runs offline on a 32 GB Mac, and it ships under Apache 2.0.
The unusual part is how it buys extra reasoning without extra size.
Inkling 2 Mini is a looped transformer. Its 24 layers are grouped into four blocks of six. The first and last blocks run once; the two middle blocks run three times per token with shared weights. Stored as 24 layers, it executes 48 layer passes. Each pass through the loop adds about 1.0B effective active parameters, so the true figure is 5.4B per token, not 3.4B.
The intuition is that middle layers do the abstract work, while the outer layers translate to and from embeddings. In the lab's own experiments, looping the middle twelve layers beat looping the last twelve on 27 of 28 evaluations.
Two consequences matter for anyone deploying it. Per-token compute is higher than the active count suggests. And setting reasoning effort to its minimum does not fully switch reasoning off.
| Spec | Value |
|---|---|
| Released | 20 Nov 2026, open weights |
| Parameters | 26.4B total, 3.4B active, 5.4B effective with looping |
| Experts | 8 of 128 per token |
| Context | Up to 1M tokens |
| Input / output | Text and images in, text out |
| Precision | Trained and served mainly in NVFP4; last four layers MXFP8 |
| Memory | About 20 GB |
| License | Apache 2.0 |
These scores are publisher-reported. On DeepSWE v1.1 it reaches 48.2%, and on Vals Index v2.2 it posts 46.5%, which the lab says leads every open model under 100B total parameters. SWE-bench Pro v2 Hard lands at 77.9% and SWE-bench Verified at 79.2%. Agentic search looks good too: 79.0% on BrowseComp and 82.7% on MCP Atlas.
Looping does not add memory for facts. Knowledge benchmarks sit where you would expect for its size, and the model leans on search and tools to compensate. It also trails Qwen3.8-27B on multi-turn chat (50.0% on MultiChallenge) and on prompts with many constraints.
Local runtime support for looped models is new, so measured speed varies a lot by runtime. Benchmark on the machine you will actually use.
Its bigger sibling, Inkling 2 Small, is the choice when quality ceiling matters more than footprint. The Mini is for laptops, edge boxes and high-throughput agent loops. The simplest trial: give both the same ten repository bug-fix tasks, with turn limits set, and compare pass rate against wall-clock time. A research variable-loop variant was released alongside it for studying how loop count affects chain-of-thought legibility.
See the full Thinking Machines lineup or other code generation models.
Compare the model on the work you need to complete.
Inkling 2 Mini is listed for Image-text-to-text, Text generation, Coding & Code Generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 1M context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
radixark
DaisyRose-27B is RadixArk's Apache 2.0, RL-only post-train of Qwen3.8-27B for agentic coding, with a 256K context and image input.
Microsoft
MAI Code 2 is Microsoft AI's coding model for multi-file work, debugging and frontend builds, live in GitHub Copilot at $0.50/$2.50 per 1M (promo).
Microsoft
MAI Code 2 Flash is Microsoft AI's low-cost coding model for edits, fixes and explanations, on every GitHub Copilot plan including Free and Student.
Microsoft
MAI Cyber 2 Flash is Microsoft AI's security model for vulnerability discovery and threat investigation, with a 1M-token context, shipping in MDASH.
Gemini 4 Hydrogen is Google's fast, lower-cost Gemini 4 model for everyday coding and agentic execution, with a 1M context, xhigh thinking and a free tier.
OpenAI
GPT-6 Cyber is OpenAI's restricted security model for authorised vulnerability investigation, reproduction and fix verification, for Daybreak Red customers.