Inkling’s open weights are served by multiple vendors, but behavior still varies
acmurthy · x · 2026-07-28
Open weights are having a moment, according to a thread that argues better models, cleaner licenses, and more ways to tune and serve are making the category more attractive.
The linked piece uses Thinking Machines’ Inkling as the example: it describes Inkling as a 975B mixture-of-experts open-weight model with an Apache 2.0 license, available across multiple vendors on launch day. The main question is whether the same weights really behave the same once they are served by different inference providers. The article compares providers on accuracy, quality, latency, and speed, and argues that vendor choice can materially change the user experience even when the underlying weights are identical.
More from Infra
- Meme contrasts a data center with bees and says no compute is worth the cost — iamaliveix · 2026-07-28
- MinerU turns PDFs and Office docs into LLM-ready Markdown to cut token costs — lxfater · 2026-07-28
- UNIUBI AI’s robot dog reportedly lands a 720° backflip on NVIDIA Orin — Scobleizer · 2026-07-28
- AI spending worries push Samsung and SK Hynix down more than 10% — nordicinst · 2026-07-28
- NVIDIA backs Ilya Sutskever’s SSI with a $5 billion strategic investment — loopuleasa · 2026-07-28
- Sentinel tests whether machine-read news can predict markets with a pre-registered live record — IPSUM99 · 2026-07-28