Laguna 118B runs on a DGX Spark, Inkling goes multimodal, K3 secures its commercial license: benchmarks and clauses recap #23

In plain terms. Three major open models released simultaneously: Laguna S2.1 (Poolside), Inkling (Thinking Machines), and Kimi K3 (Moonshot). They target different use cases—simple deployment, multimodal fine-tuning base, and maximal frontier performance—and their licenses make them distinct strategic choices for a CTO.

Context

Interconnects recap #23 (Nathan Lambert, August 2, 2026) compiles the open week. Three releases stand out, each with a distinct architecture and a licensing agreement that matters as much as the benchmarks.

The three architectures

• Laguna S2.1 (Poolside): MoE 118B / 8B active, OpenMDW license (Apache-2-like), pre- and post-training published—transparency noted by Lambert as unusual for an open release. Runs on a single DGX Spark.
• Inkling (Thinking Machines): Multimodal MoE 975B / 41B active (text, images, audio → text). Version 276B / 12B also released. Positioned by Lambert as a fine-tuning base, not a benchmark topper.
• Kimi K3 (Moonshot): 2.8T / 104B active, 1M token context, native vision (arXiv:2607.24653 - Kimi Delta Attention, Attention Residuals, Stable LatentMoE). Non-commercial license + mandatory commercial agreement—the clause sparking debate.

Analysis - who uses what

Segmentation by deployment constraints. Laguna for teams wanting minimal infra (the only one of the three executable on a single documented machine). Inkling 276B / 12B as a lightweight multimodal R&D base. K3 for maximal frontier—but the commercial agreement clause may block a US company under export constraints.

Probabilized scenarios

• K3 outside US, Laguna in Europe / US SMB, Inkling in R&D (~50%): natural segmentation by license.
• Poolside pushes Laguna in US enterprise as Fable 5 alternative (~30%).
• Closed replica of Inkling by Thinking Machines for its clients (~20%).

Implications for the pro

• Dev / ML: three solid fine-tuning bases; check K3’s commercial clause before any deployment.
• US enterprise CTO: Kimi K3 = legal arbitrage before technical.
• Data architect: Laguna, best candidate for a quick self-hosted POC.

To watch

Third-party benchmarks; first K3 implementations outside China; Inkling 276B / 12B version on Hugging Face; effective cost per served token.

So what. This week, open isn’t about one model: it’s three distinct strategic choices, to be sorted by deployment constraints and licensing terms.