Inkling: Our Open-Weights Model

Original Article ↗ Hacker News Discussion ↗

An American AI lab has released Inkling, a large open‑weights multimodal model positioned as a customizable alternative to Chinese leaders like GLM 5.2, DeepSeek, and Kimi. Commenters see it as a strong first effort—especially for instruction following and audio/vision—though it still lags top frontier and Chinese open models on many benchmarks. Much of the debate centers on whether open‑weights labs can build sustainable businesses on fine‑tuning and hosting, how much “moat” exists in a world of cheap model switching, and the strategic importance of domestically controlled open models amid growing geopolitical and regulatory pressure.

Model capabilities and benchmarks

  • Inkling is an open-weights, multimodal, long-context MoE model; there’s also a smaller 276B-total / 12B-active variant in testing.
  • Benchmarks place it below top Chinese open models like GLM‑5.2 and Kimi K2.7, and below frontier closed models (GPT‑5.x, Fable, etc.).
  • Some argue being “close” to GLM‑5.2 on a first release is impressive and shows the training moat is shrinking; others point out it debuts around #41 on one public index and question the hype.
  • Multiple commenters stress that benchmarks don’t fully capture real-world usefulness; some early private evals report the model significantly outperforming its public scores on niche tasks.

Multimodality, context, and behavior

  • Major appeal: open-weights model with vision and audio, long context (up to ~1M tokens on paper).
  • There’s debate on long-context utility: some see performance collapsing after ~150–200k tokens; others claim certain models (and tasks) still benefit well beyond that, emphasizing the need for task-specific evals.
  • Several users praise Inkling’s instruction following, tone, and “human” conversational style; coding ability is viewed as weaker than leading code-focused models.

Open-weights positioning and business models

  • Thinking Machines is seen as one of the few US labs whose incentives align with releasing strong open weights, tied to its fine‑tuning and hosting product (Tinker).
  • Proposed revenue models:
    • Hosting/inference for their own open model.
    • Managed fine‑tuning (SFT/RL) and serving custom models.
    • Being “infrastructure” for enterprises that want domain‑specific, cheaper‑than‑frontier models.
  • Many question the moat: if the weights are open, anyone (including hyperscalers) can host them; some conclude there is effectively no lasting moat and that current spending is about mindshare and domestic capability, not immediate profit.

US vs Chinese open models and geopolitics

  • Broad thread about needing “domestic” open‑weights alternatives to Chinese models (DeepSeek, GLM, Kimi, Z.ai), both in the US and elsewhere, in case of regulatory blocks or censorship.
  • Some argue Chinese models are heavily state‑constrained (e.g., Tiananmen queries), others counter that US/EU models also have political biases and content restrictions.
  • There’s concern in finance/government and similar sectors about relying on Chinese-origin models, regardless of technical quality.

Community reception

  • Many are enthusiastic: strong first release, American open‑weights, multimodal, and tightly integrated with a fine‑tuning stack.
  • Skeptics highlight: not SOTA overall, unclear cost/performance vs cheap frontier APIs, and lack of cost graphs.
  • General consensus: welcome competition, promising debut, but real value will depend on fine‑tuning workflows, pricing, and how quickly it iterates.