Latest open artifacts (#23): Laguna S2.1, Inkling, & Kimi K3 show the utility of open models on the Pareto frontier

What it claims

This roundup argues that the lab consolidation predicted for 2026–27 has not happened. More companies are training strong models, and more are releasing them openly, because demand for tokens makes "building token machines" a viable business. The page's examples are Thinking Machines, whose open-model fine-tuning service reportedly makes hundreds of millions per year, and new Chinese entrants such as Xiaomi. The author says the open questions are now how far open models can take market share, and whether revenue-share licences like Kimi K3's will stick.

How to use it

Per-model details a reader would act on, as the page gives them:

• Inkling (975B-A41B): text, image and audio in, text out. The page pitches it as a fine-tuning base, notably through Thinking Machines' commercial service Tinker, rather than as a top performer. The 276B-A12B variant is called "really competitive for its size."
• Hy3 (295B-A21B): now Apache 2.0. Its predecessor, covered in Artifacts 21, used a custom and "rather restrictive" licence, so earlier licensing objections may no longer apply.
• Laguna-S-2.1 (118B-A8B, fits on a DGX Spark): newly pre- and post-trained, not just a fine-tune. It uses the OpenMDW licence, which the page describes as Apache 2.0-like with better legal backing for AI models specifically. Poolside's blog publishes all evaluation trajectories. Laguna-XS-2.1 is a 33B-A3B update.
• DeepSeek-V4-Flash-0731: said to beat OpenAI's Luna on the price/performance Pareto frontier. It arrived one day after OpenAI cut Luna's price by 80%. The larger V4 Pro has not been updated; the page calls the original Pro "rather underwhelming."
• Kimi K3: released under a noncommercial licence. Inference and fine-tuning providers must sign a commercial agreement with Moonshot AI.
• Apertus-v1.5-70B: continued pre-training of the fully open-source Apertus 1.0 on 2T more tokens.
• Instella-MoE-16B-A3B-Think: trained on AMD Instinct cards. AMD releases every stage: base, MidTrain, SFT and DPO checkpoints. This is useful for anyone studying or reproducing the post-training stages.
• Motif-3-Beta (314B-A13B): a preview that introduces the architectural components GDLA and mHC. The page does not explain them.
• LongCat-2.0 (1.6T MoE): trained entirely on Huawei Ascend 910s. The page calls it the first non-Huawei, non-toy model trained fully on Chinese accelerators, noting that other Chinese chips are mostly used for inference, if at all.

Limits

• Kimi K3 carries legal exposure. The page cites Kevin Xu and Graham Webster: if a US company needs a contract with Moonshot to serve K3 tokens, policy levers debated for restricting Chinese open-model use "would more clearly apply." US inference providers should check this before adopting.
• Hy3's math result is lightly supported. The claimed proof of a 50-year-old math problem used a dedicated harness, with the model Sol acting as judge. The author says it is unclear how much the judge mattered.
• Several models are flagged as weaker than their headline suggests:
• Inkling is "not the strongest model among peers (in China) in its size class."
• LongCat-2.0 is "not the most capable for its size beyond benchmarks."
• Motif-3-Beta is explicitly a preview.
• The page gives no benchmark numbers. It states no scores, baselines or settings for any model, so every performance claim above is the author's characterisation only.