Inkling 2 Small Free

Thinking Machines Lab released Inkling 2 Small on November 5, 2026, a sparse Mixture-of-Experts model with 594B total parameters, 17B active per token, a context window up to 1 million tokens, and text plus image input producing text output. Unlike the July Inkling-Small, whose 276B/12B weights were published at launch, Inkling 2 Small ships closed-weight, with Apache 2.0 weights promised the following week once safety testing and red-teaming finish. The model card describes a 48-layer decoder-only transformer routing each token to 8 of 384 experts with hybrid local/global attention, plus a 292M-parameter Perception Encoder-based vision encoder using 20x20-pixel patches and 2x2 pooling. Benchmark trajectories ran at a 1M-token limit, up from 256K for Inkling-Small, while reinforcement learning used a 256K window with compaction; the reported Artificial Analysis Intelligence Index score is 30.6.

An ArtificialWatch radar sweep recorded the exact catalog id opencode/inkling-2-small-free appearing in models.dev on Nov 7, 2026 at a listed price of $0 per 1M tokens with a 1,048,576-token context window. The entry is explicitly machine-written from the sweep and states it proves only that the id now exists at that price and context size. The same registry note traces Inkling 2 Small to Thinking Machines' Nov 5 launch on Tinker with closed weights, describing a 594B-total/17B-active MoE, image input, and variable effort, with full weights promised under Apache 2.0 after safety testing. It reports an Artificial Analysis Intelligence Index score of 31 and OpenCode having hosted early checkpoints as Fledge-Alpha.

Artificial Analysis reports Thinking Machines released Inkling 2 Small on November 5, 2026, scoring 31 on its Intelligence Index, up 18 points from Inkling Small with about 4x as many output tokens per task. If the planned weights release happens, the model would become the highest-scoring open-weights model developed outside China. It matches Google DeepMind's Gemini 3.8 Flash (high) and surpasses Xiaomi's MiMo-V2.6-Pro on the Intelligence Index v5.0.1. Benchmark shifts are concentrated in agentic work: AutomationBench-AA rises from 5% to 46%, FrontierCode 2.0 from 5% to 25%, and HLE-Diamond from 12% to 29%. Coding profile differs from Claude Sonnet 5.5 (medium) at 25% versus 11% on FrontierCode 2.0 but 13% versus 30% on Terminal-Bench 4.0, and hallucination rate on AA-Omniscience falls from 61% to 27%.

Thinking Machines Lab lists Inkling 2 Small as the newest entry in its Inkling family, billed as its most capable model and the first of a second generation, with a large jump in agentic coding and vision. The model card gives a sparse mixture-of-experts profile of 594B total parameters with 17B active per token, alongside the larger 975B/41B Inkling and the 276B/12B Inkling-Small. The vendor page states that Inkling 2 Small ships closed-weight at release, with full weights intended for release in coming weeks pending additional safety testing, and it is available for inference and fine-tuning on the company's Tinker training platform. Highlighted capabilities include general intelligence across knowledge, math and science, agentic coding and tool use, controllable reasoning effort, calibrated forecasting confidence, and reliable multi-step instruction following.

A November 8, 2026 entry in the IADecider AI news archive reports that OpenCode's model registry now lists the model ID inkling-2-small-free, marked "free for a limited time" and flagged as having closed weights, served on an OpenAI-compatible inference endpoint. The listing covers text and image input with a one-million-token context, 131K max output, and adjustable reasoning effort. Pricing is listed as a flat $0 per million tokens both in and out, giving inference-first testers a no-cost way to try the model before weights are published. The same entry notes that Thinking Machines had not announced the arrangement on its own channels, so the free route should be read as an OpenCode offer rather than a Tinker product, and the weights claim may already have moved on since November 8.

This tracker page records Inkling 2 Small as API-only with weights still pending, advertising a 1M-token context while served context variants are 64K and 256K, with an Intelligence Index score of 31 and an output price of $1 per million tokens. Full weights are promised under Apache 2.0 around November 12, 2026, and the model is currently reachable for inference and fine-tuning on Tinker under Tinker Terms of Use. The tracker lists architecture details matching the model card: 48 layers, sparse MoE routing each token to 8 of 384 experts, hybrid local/global attention, and native image input through a 292M-parameter vision encoder with 20x20-pixel patches and 2x2 spatial pooling. Its 71-row coding comparison table places the model mid-pack on most benchmarks, with clearer wins in finance and visual design tasks.

Blockchain.News' AI Radar reports that Thinking Machines made the Inkling 2 Small model available for free on OpenCode for a limited time in order to gather improvements, citing a post by Soumith Chintala on Twitter/X dated Nov 7, 2026. The item is one line in a multi-story aggregation feed. The same feed covers adjacent AI developer news, including DeepLearning.AI hosting a talk on giving coding agents repository and decision context at AI Dev in New York on Nov 30–Dec 1, 2026, and an Anthropic engineering blog post describing a dreaming process that helps Claude organize learned information and retrieve knowledge more efficiently.