1. VEGA
  2. Blog
  3. Inkling 2 Small: Thinking Machines' Best Open Model Yet
November 13, 2026Announcements

Inkling 2 Small: Thinking Machines' Best Open Model Yet

Inkling 2 Small: Thinking Machines' Best Open Model Yet

Thinking Machines Lab has released Inkling 2 Small, its first second-generation model and, at launch, the strongest open-weights release built outside China on the Artificial Analysis Intelligence Index. It combines agentic coding, image understanding, and a 1M-token context window with Apache 2.0 weights. For readers choosing a model for real work, the main draw is its ability to make changes, check the results, and keep working rather than produce a convincing first answer and stop.

What is Inkling 2 Small?

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, announced Inkling 2 Small on November 5, 2026, through Tinker and Tinker Playground. Its open weights followed on Hugging Face on November 12 under Apache 2.0.

Despite the name, this is not a small model to host. Inkling 2 Small is a sparse Mixture-of-Experts model with 594 billion total parameters and approximately 17 billion active per token. It selects eight of its 384 experts for each token.

Its context window holds 1,048,576 tokens, giving it room for long documents and substantial code context. Inputs can include text and images; unlike the first Inkling, it does not accept audio.

Developers can also adjust reasoning_effort from none to max, balancing reasoning depth against cost. In practice that means cheap, quick answers for simple questions and deeper work when a problem earns it.

Inkling 2 Small benchmarks: the biggest gains

The headline result is 31 points on Artificial Analysis Intelligence Index v5.0.1, up from 13 for Inkling Small. That places it above every open model built outside China in that evaluation, at a level comparable to Gemini 3.8 Flash with high reasoning effort.

Several results make the improvement more concrete:

BenchmarkInkling 2 Small resultWhy it matters
SWE-Bench Pro v2 (Hard)86.3%Strong performance on difficult coding tasks
ARC-AGI-281.7%Best open-weights result, according to ARC Prize
AA Intelligence Index v5.0.131, versus 13 previouslyA substantial generation-over-generation gain
AA-Briefcase1527 Elo, versus 912 previouslyBetter performance on long business tasks
AA-Omniscience v1.127% hallucination share, down from 61%More selective answering reduced hallucinations

On AA-Briefcase, which covers spreadsheets, presentations, and memos, its 1527 Elo also exceeds Claude Sonnet 5.5 at medium reasoning effort, with 1439, and Gemini 3.8 Flash at high effort, with 1203.

The hallucination result needs careful interpretation. The reduction came mostly through more selective answering, not the elimination of factual errors. It is an improvement, but important claims still need checking.

What Inkling 2 Small is good for in practice

Agentic coding is the clearest starting point. Training included hundreds of thousands of long coding tasks, with an emphasis on following instructions, evaluating changes, and checking completed work. That makes repository tasks a more revealing test than asking for an isolated function.

Image input adds a useful capability: vision-in-the-loop self-verification. An agent can generate an e-commerce site, inspect its rendered appearance, and use that visual feedback to assess its own output. This goes beyond describing a supplied picture.

For finance and knowledge work, the AA-Briefcase results suggest testing it on assignments with several deliverables rather than a single summary. It also scored 67.5% on Finance Agent v2.1.

Data analysis is another promising use case. On Plotly DataAnalyticsBench, it scored 39 out of 43, or 91%, with perfect scores in categories covering ambiguity, infeasible requests, data-quality problems, and statistical traps.

For research workflows, its training included "autoresearch" exercises such as fine-tuning neural networks and optimizing GPU kernels. In one reported example, it independently fine-tuned Qwen3.5-4B for time-series forecasting, doubling that model's score in 4.5 hours for $426 on Tinker. That is a specific demonstration, not a guarantee for other experiments.

Apache 2.0 openness and the Inkling 2 Small ecosystem

Apache 2.0 weights make Inkling 2 Small available for customization, including enterprise ownership of fine-tuned versions. Thinking Machines also reports faster fine-tuning and higher resulting quality than its previous models when using identical data and algorithms.

Day-zero support included vLLM, SGLang, Together AI, Snowflake Cortex, and Cloudflare Workers AI. The model is also available through OpenRouter.

Serving optimizations matter at this scale. DSpark speculative decoding drafts seven tokens per step and reached up to approximately 630 tokens per second per user at TP4 on GB200. SGLang demonstrated 576 tokens per second on NVIDIA Vera Rubin. These are specific hardware results, not universal speed promises.

Local deployment remains demanding: a four-bit quantized version needs approximately 356 GB of memory.

Where Inkling 2 Small still struggles

Results are uneven. Inkling 2 Small scored just 13.1% on Terminal-Bench 4.0, and Artificial Analysis measured approximately 21.5 minutes per Intelligence Index task. Long tasks can therefore be slow despite high token-generation speeds. Known limitations also include hallucinations and degradation during long multi-turn conversations; a large context window does not remove those weaknesses.

Try Inkling 2 Small on vega.chat

Inkling 2 Small is already live in the vega.chat model picker, under both the chat and code groups. Image input is supported; PDF input is not. Give it a real coding task or a long document next to your usual model, and see how it holds up when it has to check its own work.

Developers can access it through api.vega.chat, the OpenRouter-compatible API, to test it in agents, coding workflows, and their own applications.

VEGA API — all models·API docs·Inkling 2 Small: лучшая открытая модель Thinking Machines

Read also

Qwen 4 Max: a 2.4T-parameter teammate for long-running work
Qwen 4 Max: a 2.4T-parameter teammate for long-running workNovember 16, 2026
Announcing Gemini 4 Hydrogen
Announcing Gemini 4 HydrogenNovember 13, 2026
MiniMax H3-Image: Open-Weight 4K Images in Under 10 Seconds
MiniMax H3-Image: Open-Weight 4K Images in Under 10 SecondsNovember 11, 2026

Comments · 0

VEGA - built in getmyai.io
PricingModelsBlogAbout CreatorTerms of ServicePrivacy PolicyMailing Policy