Thinking Machines releases Inkling 2 Small with Apache 2.0 plan
Thinking Machines Lab has released Inkling 2 Small, the first model in its second Inkling generation. It is a sparse Mixture of Experts transformer with 594 billion total parameters and 17 billion active parameters. The model accepts text and images and supports variable reasoning effort. It runs on Tinker for fine-tuning and inference today.
The release post reports a score of 31 on the Artificial Analysis Intelligence Index v5.0.1, up from 13 for the previous Inkling Small. Weights are closed at launch, but Thinking Machines Lab plans an Apache 2.0 release after external safety testing and red-teaming. Separately, Vals AI researchers read internal thinking logs from 2,655 benchmark runs and found profanity in 50 of them.
In one line: Thinking Machines Lab ships Inkling 2 Small, a 594 billion parameter sparse MoE with 17 billion active parameters, a 31 on the AA Intelligence Index, and Apache 2.0 weights planned after safety checks.
Model architecture and training
Inkling 2 Small is a 48 layer decoder-only transformer. Each token routes to 8 of 384 experts. Attention mixes local and global layers. A 292 million parameter vision encoder, based on Perception Encoder, uses 20 by 20 pixel patches with 2 by 2 spatial pooling. Image patches are interleaved with text tokens. Weights support MXFP8 and NVFP4, MarkTechPost reports.
Pretraining ran in four phases. H1 used 8.4 trillion tokens over 68 hours on NVIDIA GB300 NVL72 racks, trained natively in NVFP4. H2 used 3.8 trillion tokens of newer, higher quality data at 512K context. H3 used about 320 billion tokens of reasoning and agent trajectories at 1 million token context. H4 used about 34 billion tokens of the highest quality agent trajectories.
A single joint reinforcement learning run followed. It took 1,075 steps, more than 17 million rollouts and 9 days on 2,048 GPUs. Compute exceeded 4e23 FLOPs. Coding made up 65 percent of RL tokens. Alignment included an explicit reward and on-policy self-distillation from a conduct spec. Thinking Machines Lab says its internal misbehavior score fell from 66.8 on the previous Inkling to 14.9.
Benchmarks show coding strengths and terminal gaps
All task scores below come from the lab's benchmark table, run at effort 0.99 and temperature 1.0. Most rows use third-party evaluators such as Artificial Analysis, Scale AI and Vals.
Where it scores highest:
- SWE-Bench Pro v2 Hard: 86.3 percent.
- PerceptionBench: 64.7 percent, versus 58.5 percent for Kimi K3.
- EnterpriseBench-CoreCraft: 71.8 percent, versus 46.7 percent for GLM-5.3-Flash.
- JobBench: 70.5 percent.
- DesignArena Website: 1,340, third on the November 5 leaderboard snapshot.
Where it trails:
- Terminal-Bench 4.0: 13.1 percent, versus 33.8 percent for DeepSeek-V4.1-Flash-1017.
- DeepSWE v1.1: 72.1 percent, versus 77.0 percent for DeepSeek-V4.1-Flash-1017.
- GDPVal-AA v2.1: 1,630, versus 1,746 for DeepSeek-V4.1-Flash-1017.
On the PMPP-Hard GPU kernel benchmark, it solved 54.25 of 69 tasks on average. In an autoresearch demo, it fine-tuned Qwen3.5-4B for time-series forecasting and spent $426 of a $500 Tinker budget. The post cites both 4.5 hours and 4 hours 17 minutes for that run.
Artificial Analysis lists pricing at $0.30 per 1 million input tokens and $1.00 per 1 million output tokens. The index cost per task is $1.18 and time per task is 21.5 minutes. DeepSeek-V4.1-Flash-1017 scores 33, costs $0.60 per task and takes 13.3 minutes. GLM-5.3-Flash scores 28, costs $0.58 and takes 37.7 minutes.
Swearing in chain of thought
Vals AI read the model's internal thinking logs from 2,655 runs across 13 benchmarks. It found "fuck" or variations in 50 runs, totalling 316 instances. 241 came from just two runs that lost control. In one, the model was asked to calculate an asteroid trajectory and produced a single 500,000 character response, roughly novel length, repeating the same sentences with 158 expletives.
ProgramBench was hit hardest, with 20 of 200 runs affected. The profanity usually appears when the model makes mistakes, thinks it has a breakthrough or gives up. One run read: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS."
The swearing does not work as a warning signal. Vals AI reports that 46 of 48 scored runs with swearing failed. But about 75 percent of all scored runs fail anyway, and 96 percent of failed runs contained no expletives. Users typically do not see these thinking steps, or see them only with a filter.
What this means for small delivery teams
For a team that ships software with agentic coding tools, the coding tools directory is a useful place to compare alternatives, and my process for shipping client products without a dev team shows how I evaluate model outputs on real delivery work. Inkling 2 Small's SWE-Bench Pro v2 Hard score of 86.3 percent is the kind of number you notice. PerceptionBench and EnterpriseBench-CoreCraft also beat direct rivals in the lab's table. But Terminal-Bench 4.0 at 13.1 percent says this model still struggles with live terminal tasks, and the chain of thought instability is a real concern if you expose raw reasoning to users.
You can try it today on Tinker for fine-tuning and inference. If you have a client project that needs a RAG chatbot or an internal AI rollout, the price per output token is cheaper than the previous Inkling Small. But weights are not out yet, and self-hosting hardware is not disclosed. I would watch the planned Apache 2.0 release and test the model against the same cases you would use for any model. The model is not a drop-in replacement for DeepSeek or GLM on every task.
Quick answers
What is Inkling 2 Small?
Inkling 2 Small is Thinking Machines Lab's 594B total parameter MoE model with 17B active parameters, text and image input, and a 1M token context. It runs on Tinker and scores 31 on the Artificial Analysis Intelligence Index.
Is Inkling 2 Small open weights?
Not at launch. Thinking Machines Lab says it plans an Apache 2.0 release after external safety testing and red-teaming, following its safe path to open weights policy. Checkpoints become downloadable when weights ship.
Why did Inkling 2 Small swear in benchmark runs?
Vals AI found 316 instances of profanity in internal chain-of-thought logs across 50 of 2,655 runs. Thinking Machines has not said it fixed this; users typically do not see raw thinking steps or see them only with a filter.