Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026, a sparse mixture-of-experts model that scored 31 on the Artificial Analysis Intelligence Index v5.0.1, matching Google DeepMind's Gemini 3.8 Flash (high) and placing it at the top of the index among models built outside China2. The company plans to release full weights under an Apache 2.0 license after external safety testing and red-teaming, which would make it the highest-scoring open-weights model developed outside China on current rankings, according to Artificial Analysis1.
Inkling 2 Small is launching as a proprietary model, available first through Thinking Machines Lab's Tinker platform and Tinker Playground3. The model is a 594-billion-parameter system with 17 billion active parameters, built on a 48-layer decoder-only transformer that routes each token to eight of 384 experts. It accepts text and image input, generates text output, and supports a 1-million-token context window with MXFP8 and NVFP4 numerics.
Training and architecture
Initial pretraining covered 8.4 trillion tokens of natural-language, code, and image data over 68 hours on NVIDIA GB300 NVL72 racks using native NVFP4 training. Unlike the original Inkling, which launched with full weights from day one, Inkling 2 Small starts closed, with Tinker Playground providing inference-only access and Tinker supporting fine-tuning and sampling. Thinking Machines Lab intends to publish weights in the week following the announcement, conditional on completed safety testing.
Benchmark gains over predecessor
The model's index score is up 18 points from its predecessor, Inkling Small, while generating roughly four times as many output tokens per task. The sharpest gains appear on agentic and long-horizon work: AutomationBench-AA rises from 5% to 46%, FrontierCode 2.0 climbs from 5% to 25%, and HLE-Diamond moves from 12% to 29%. On CritPt v2, the model scores 56%, up from Inkling Small's 34%. Inkling 2 Small also edges out Gemini 3.8 Flash on Code2Video-AA and AA-3D, though it trails Claude Sonnet 5.5 on both benchmarks.
On the AA-Briefcase evaluation, Inkling 2 Small scores 1,527 Elo, clearing Gemini 3.8 Flash (high) at 1,203 Elo and Claude Sonnet 5.5 (medium) at 1,439 Elo, up from Inkling Small's 912 Elo. On FrontierCode 2.0, Inkling 2 Small's 25% score exceeds Claude Sonnet 5.5 (medium) at 11%.
Pricing and cost position
Token pricing is set at $0.30 per million input tokens and $1.00 per million output tokens, with cached input tokens at $0.03 per million. Per Intelligence Index task, Inkling 2 Small costs $1.18, compared with $7.77 for Gemini 3.8 Flash (high) and $10.02 for Claude Sonnet 5.5 (medium). The model averages around 150,000 output tokens per task. Inkling Small cost almost six times less per task than Inkling 2 Small, reflecting the new model's heavier token generation.
On knowledge accuracy, Inkling 2 Small's hallucination rate on AA-Omniscience v1.1 drops from 61% to 27% versus its predecessor, while accuracy edges from 32% to 34% and the attempt rate falls from 77% to 53%.
ANALYSIS The pricing gap is notable: Inkling 2 Small delivers index-comparable performance to Gemini 3.8 Flash (high) at roughly one-sixth the per-task cost, a ratio that could shift once the weights are open and third-party inference providers compete on serving.
The planned Apache 2.0 weight release arrives as other labs also push large open-weights models; Moonshot AI published the complete weights for its 2.8-trillion-parameter Kimi K3.1 Pro on November 4[1]. Thinking Machines Lab's weight release timeline is conditional on safety testing completion.