Open Model Race Intensifies: Thinking Machines, Tencent, Poolside, Moonshot Ship Frontier-Class Releases in Same Week

The consolidation thesis is failing

For years, the consensus prediction was that rising training costs would force consolidation among AI labs — fewer companies, bigger checks, higher barriers to entry. According to Interconnects.ai's latest "Open Artifacts" roundup, that thesis is not playing out. Instead, more organizations are training frontier-class models and releasing them openly, even as per-model training costs climb into the hundreds of millions of dollars.

The clearest evidence: Thinking Machines, founded in February 2025, has gone from an unproven startup to what Interconnects.ai describes as the top U.S. open-weight model producer — ahead of established players like NVIDIA (Nemotron) and Arcee (Trilogy). The company's Tinker fine-tuning service reportedly generates hundreds of millions in annual revenue, according to the report.

The releases

Inkling (Thinking Machines): A 975B-total/41B-active parameter multimodal Mixture-of-Experts (MoE) model accepting text, image, and audio inputs and producing text output. It is not the strongest model in its size class relative to Chinese competitors, but Interconnects.ai notes it is designed as a fine-tuning base for Thinking Machines' Tinker platform. A smaller 276B-A12B variant is also competitive for its size class.

Hy3 (Tencent): A 295B-A21B MoE that improves on its predecessor (Hunyuan, covered in Artifacts #21) across all measured metrics. The larger change is licensing: Tencent moved from a restrictive custom license to Apache 2.0. The model reportedly proved a 50-year-old math problem using a dedicated harness with Sol as judge — though the report cautions the judge's role in that result is unclear.

Laguna-S-2.1 (Poolside): A 118B-A8B MoE, small enough to run on an NVIDIA DGX Spark, drawing significant attention. Poolside adopted the OpenMDW license, an Apache 2.0-style license built with stronger legal backing for AI models. This marks Poolside's third consecutive monthly appearance in the roundup, alongside a smaller Laguna-XS-2.1 update (33B-A3B).

DeepSeek-V4-Flash-0731: Released one day after OpenAI cut pricing on its smallest model by 80%, this update reportedly beats rival "Luna" on the pareto frontier of cost versus performance. DeepSeek's larger V4-Pro model has not yet received the same update.

Kimi-K3 (Moonshot AI): Described as the largest open model release in recent memory, Kimi K3 ships under a noncommercial license requiring inference and fine-tuning providers to sign a separate commercial agreement. Commentators Kevin Xu and Graham Webster argue this licensing structure could expose U.S. companies using Kimi K3 tokens to future government restrictions on doing business with Chinese AI firms.

LongCat-2.0 (Meituan): A 1.6 trillion-parameter MoE trained entirely on Huawei Ascend 910 accelerators — reportedly the first non-Huawei, non-toy model trained fully on Chinese chips rather than using them only for inference.

Other releases include Motif-3-Beta (314B-A13B, Motif Technologies, South Korea) with new GDLA and mHC architectural components, Apertus-v1.5-70B (Swiss AI) trained on 2 trillion additional tokens, and Instella-MoE-16B-A3B-Think (AMD), trained on Instinct accelerators with full checkpoint releases across pretraining, SFT, and DPO stages.

What this means

The volume and diversity of this release cycle undercuts the consolidation narrative that has dominated AI industry forecasting since 2023. Instead of a shrinking set of labs with proprietary frontier models, capacity to train competitive models is spreading across the U.S., China, Switzerland, and South Korea. The more consequential trend may be licensing: Moonshot's noncommercial terms on Kimi K3 and the shift toward permissive licenses like OpenMDW and Apache 2.0 suggest legal structure — not just model quality — will determine which open models see real commercial adoption, and which become entangled in U.S.-China trade policy.