Thinking Machines' Inkling 2 Small: Scores 31 on Artificial Analysis, Ships Closed, Promises Apache 2.0 Weights Next Week
The 594-billion-parameter model beats the lab's first generation but trails Chinese open models on terminal and kernel-optimization benchmarks in its own tables. The open license is a promise, not a fact, until the weights appear.
TL;DR
Thinking Machines Lab released Inkling 2 Small on 5 November, a mixture-of-experts model with 594 billion total and 17 billion active parameters that takes image input. It scores 31 on the Artificial Analysis Intelligence Index and is available through the lab's Tinker platform with closed weights. The company says it intends to publish the weights under Apache 2.0 next week, after external safety testing and red-teaming. Its own model card shows it behind DeepSeek V4.1 Flash and GLM 5.3 on several coding and terminal benchmarks.
Thinking Machines Lab introduced Inkling 2 Small on 5 November as its second model generation. It is a mixture-of-experts model with 594 billion total parameters and 17 billion active, accepts text and images, and has variable effort settings. The lab says the gains over its first generation come from more efficient pre-training and a large, unified reinforcement-learning run.
What you can use today, and what's still a promise
| Question | Answer | Source |
|---|---|---|
| Where is it available? | Tinker Playground, and Tinker for fine-tuning and sampling | Launch post |
| Are the weights out? | No. Closed for now | Launch post |
| What is promised? | Full weights under Apache 2.0 "next week", after external safety testing and red-teaming. Checkpoints become downloadable at that point | Launch post |
| License today | Tinker Terms of Use, per the model card; llm-stats lists it as proprietary | Model card, llm-stats |
| Context window | Up to 1M tokens per the model card; llm-stats lists 256K for the hosted endpoint | Model card, llm-stats |
| Price | $0.30 per million input tokens, $1.00 output, $0.03 cached input, per llm-stats | llm-stats |
The price and the 256K figure come from llm-stats, not from Thinking Machines, and we couldn't confirm them on Tinker's own pages. RuntimeWire notes the release is a test of whether the lab can repeat a cycle of shipping through Tinker and then releasing open weights.
How it scores
| Benchmark | Inkling 2 Small | GLM-5.3-Flash | DeepSeek-V4.1-Flash-1017 | Claude Haiku 5.5 |
|---|---|---|---|---|
| Artificial Analysis Index (v5.0) | 30.6 | 27.8 | 32.6 | 36.5 |
| DeepSWE v1.1 | 72.1% | 63.4% | 77.0% | 59.9% |
| Terminal Bench v4.0 | 13.1% | 32.8% | 33.8% | 32.8% |
| Vibe Code Bench v1.1 | 81.6% | 30.8% | 87.7% | 90.4% |
| FrontierCode v2.0 | 25.4% | 14.2% | 28.5% | 34.9% |
- The launch post rounds the Artificial Analysis score to 31 and says it is far ahead of the lab's earlier models. The model card shows 30.6.
- It does well against the smaller Flash-class open models on DeepSWE and Vibe Code Bench, and poorly on Terminal Bench v4.0, where its 13.1% is well under GLM-5.3-Flash's 32.8%.
- On the card's kernel-optimization benchmark (PMPP-Hard, 69 tasks, a 10-hour budget), it solved a mean of 54.25 tasks, behind GPT-6 Astra at 63, Kimi K3.1 Pro at 60.5 and Claude Opus 5.5 at 60.
What Thinking Machines is claiming
The launch post says the model was trained with reinforcement learning on hundreds of thousands of long-horizon coding tasks across several agent harnesses. Many tasks used an agentic grader that judged how the model approached the problem, not only whether tests passed. Early checkpoints were tested on OpenCode under the name Fledge-Alpha. The lab says the model follows instructions better and verifies its work without being asked. It shows demos built with it, such as a Three.js city simulator and a Python PC emulator that boots a Linux kernel. These are the lab's own examples.
What this means if you're choosing a model
- Don't plan around the open weights yet. The Apache 2.0 release depends on safety testing finishing, and until then the model is a hosted API with its own terms.
- It's a better fit for fine-tuning than as a drop-in agent. The strongest part of the offer is Tinker's fine-tuning, and the weakest numbers are terminal tasks.
- Check the license on release day. The model card already points to a separate Model Acceptable Use Policy, so read every term that applies to the weights, not just the license name.
Sources
AI Industry Reporter
Priya covers model releases, industry announcements, and the gap between what labs claim and what independent evaluators actually find. She reads the primary source - the paper, the system card, the benchmark org's own statement - before writing a word.
More on AI Chatbots & Assistants
Microsoft Starts Capping Free Copilot Chat for Companies With 2,000+ Seats, and Won't Say Where the Cap Is
Priya Nair · 4 min
Anthropic's Amended IPO Filing: $16.96 Billion in Q3 Revenue, a $567 Million Operating Loss and $5 Billion of Short-Term Compute
Priya Nair · 5 min
Ai2's StateBench: Frontier Models Fall to 13% or Lower When Long Documents Keep Changing
Priya Nair · 4 min