# Thinking Machines Lab releases Inkling 2 Small, a 594-billion-parameter MoE priced at $0.30 per million input tokens

The model activates 17 billion parameters per token and reads up to 1 million tokens of text and images. The lab says Apache 2.0 weights will follow next week, and the license would carry none of the revenue conditions attached to Moonshot's newly published Kimi K3.1 Pro.

- Published: 2026-11-07T06:25:02.319Z
- Canonical: https://polylog.news/ai/2026-11-07/thinking-machines-lab-releases-inkling-2-small-a-594-billion
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Thinking Machines Lab](https://thinkingmachines.ai/news/inkling-2-small/), [Latent Space](https://latent.space/p/ainews-inkling-2-small-17a-594b-trained-in-68-hours-on-12t-tokens-beating-muse-spark/), [MarkTechPost](https://www.marktechpost.com/2026/11/05/thinking-machines-inkling-2-small-release/), [Pasquale Pillitteri](https://pasqualepillitteri.it/en/news/30515/inkling-2-small-thinking-machines-launch-open-weights)

Thinking Machines Lab, the company founded by former OpenAI chief technology officer Mira Murati, [released Inkling 2 Small](https://thinkingmachines.ai/news/inkling-2-small/) on November 5. It is a MoE model with 594 billion total parameters and 17 billion active. It has a 292-million-parameter vision encoder based on Perception Encoder, variable reasoning effort, and a 1-million-token context window, according to the [lab and secondary summaries](https://cryptobriefing.com/thinking-machines-inkling-2-small-release/). The Russian-language channel AI ML Big Data [described it](https://t.me/ai_machinelearning_big_data/11301) as a multimodal release that took third place on DesignArena.

For now the model runs only on Tinker, the lab's fine-tuning and inference service. Reported prices are $0.30 per million input tokens, $0.03 for cached input and $1.00 for output, with training priced at $1.50 per million tokens at 64K context ([Cellcog](https://cellcog.ai/blog/inkling-2-small/)). Meta charged $1.25 per million input tokens and $4.25 per million output tokens for its first paid Muse Spark API ([AI Weekly](https://aiweekly.co/alerts/meta-prices-muse-spark-11-api-at-125425-per-m-tokens)). The lab says Apache 2.0 weights will follow next week after safety testing. Until they ship, [Artificial Analysis lists the model as proprietary](https://latent.space/p/ainews-inkling-2-small-17a-594b-trained-in-68-hours-on-12t-tokens-beating-muse-spark/).

The training disclosures are unusually specific. The lab says initial pretraining covered 8.4 trillion tokens of text, code and images in 68 hours on NVIDIA GB300 NVL72 racks. It says the main RL run used 2,048 GPUs for nine days, with more than 17 million rollouts, more than 3 trillion tokens and more than 4×10²³ FLOPs ([Latent Space](https://latent.space/p/ainews-inkling-2-small-17a-594b-trained-in-68-hours-on-12t-tokens-beating-muse-spark/)). Under the standard 6ND approximation with 17 billion active parameters, the desk estimates initial pretraining at about 8.6×10²³ FLOPs. On those numbers the RL stage used roughly half as much compute as pretraining. Secondary tallies that add mid-training data at 512K context reach about 12 trillion tokens. That total is not the lab's headline figure.

All the benchmark numbers come from the vendor and have not been reproduced independently. The lab reports 86.3% on SWE-Bench Pro v2 (Hard), second to Moonshot's Kimi K3 at 88.2% ([Pillitteri](https://pasqualepillitteri.it/en/news/30515/inkling-2-small-thinking-machines-launch-open-weights)). Agentic terminal work is the weak spot. On Terminal-Bench 4.0 the model scores 13.1%, against 33.8% for DeepSeek-V4.1-Flash and 32.8% for GLM-5.3-Flash. On Terminal-Bench 2.1 the lab reports 85.1% with a different harness, against 90.6% for DeepSeek-V4.1-Flash ([MarkTechPost](https://www.marktechpost.com/2026/11/05/thinking-machines-inkling-2-small-release/)). The model scores 31 on the Artificial Analysis Intelligence Index. Sources disagree on its DesignArena rank: Latent Space puts it fourth overall and first among open-weight models.

The day before, Moonshot AI [published the Kimi K3.1 Pro weights](https://www.aimodeling.com/en/news/slug/kimi-k3-1-pro-open-weights-huggingface) on Hugging Face. That model has 2.8 trillion total and 104 billion active parameters in MXFP4, takes about 1.56 terabytes on disk, and ships under a custom license with a reported revenue condition for large hosted services ([DEV Community](https://dev.to/yong_yu_f98e15562e9b120a0/kimi-k31-pro-weights-are-out-20m-license-cliff-156-tb-reality-and-hosted-vs-self-host-2ie4)). The two releases give buyers a clear choice. Inkling is smaller, cheaper to serve and comes from a US lab. Kimi is larger and stronger on several agent benchmarks, and it comes from a Chinese lab.

## What this means

Inkling 2 Small competes on three things at once: a permissive license, a low active-parameter count, and US origin. Those matter most to regulated Western buyers who want weights they can download but will not deploy Chinese ones. That puts pressure on two groups. One is the Chinese open-weight vendors (Moonshot, DeepSeek, Zhipu) selling into those buyers. The other is the mid-tier closed APIs priced several times higher. The Terminal-Bench 4.0 gap shows the model is not yet a frontier agent, so it supports the open-weight trend for coding and works against it for long-horizon terminal tasks. The closest precedent is Meta's Llama 2 in July 2023. Its commercial license led to widespread enterprise fine-tuning, and Meta lost that position once Qwen and DeepSeek beat it on capability. The desk's forward read is that if the Apache 2.0 weights ship within the promised week without new restrictions, US hosting providers will list the model within weeks. Within months, at least one closed vendor should cut mid-tier prices toward the $1-per-million-output level. That would push API margins down and lift demand for GPUs used to self-host. The read is wrong if the weights slip past the stated window or arrive with usage limits, or if the weak terminal-agent scores keep enterprises on Kimi and DeepSeek for agent work.

## What to watch

- Whether the Apache 2.0 weights appear on Hugging Face next week as promised. A delay or a narrower license would mean Inkling is a hosted product, not a US open-weight alternative.
- Independent reruns of SWE-Bench Pro v2 and Terminal-Bench 4.0 on the released weights, which will show whether the vendor numbers and the terminal-agent weakness hold up.
- Price changes on mid-tier closed APIs from Meta, Google and OpenAI in the coming weeks, which would show whether $0.30 and $1.00 per million tokens is becoming the reference price.