Open-Weight Models

Definition

• 2026-08-03: alibaba commits to open-weight qwen3-8-max (and qwen3-8-27b) “next week” — not yet released (2026-08-03-alibaba-qwen38-max-alicloud-press)

AI models whose trained weights are publicly released, enabling self-hosting, fine-tuning, and sovereign deployment without API dependency.

Key Points

• 2026-07-27: open-secure-ai-alliance frames open weights as cyber-defense assets post-HF incident (2026-07-27-nvidia-open-secure-ai-alliance-blog)

• 2026-07-27: Scheduled kimi-k3 full weights (largest open-class claim); self-host vs API sovereignty (2026-07-26-kimi-k3-open-weights-july-27-release)

• 2026-07: kat-coder-v2-5-Dev Apache-2.0 open weights on HF (separate from served Pro) (2026-07-26-kwaikat-kat-coder-v2-5)

• 2026-07-21: cisco-antares 350M/1B open-weight (gated defenders) for vulnerability-localization (2026-07-23-cisco-antares-open-weight-vuln-localization)

• 2026-07: hugging-face used self-hosted glm-5-2 for IR after commercial APIs blocked exploit-payload analysis (hugging-face-ai-agent-security-incident)

• 2026-07-19: qwen-38 promises open weights “soon” while Max Preview is hosted; would break Alibaba pattern of keeping Max-tier closed (2026-07-19-alibaba-qwen-38-officechai)

• 2026-07-16: nemotron-3-embed open embedding collection (OpenMDW-1.1) tops rteb (nvidia-nemotron-3-embed-rteb)

• 2026-07-16: kimi-k3 claims largest open-class model (2.8T); weights Jul 27 — hosted API proprietary until then (2026-07-17-moonshot-kimi-k3-trilogyai)

• 2026-07-15: thinking-machines-lab releases inkling (975B/41B MoE, Apache 2.0) — largest U.S. open-weights model to date; HF + tinker (2026-07-16-thinking-machines-inkling-official)

• 2026-07-09: ollama raises $65M Series B — 8.9M monthly developers, 85% Fortune 500; largest open-model developer network (2026-07-09-ollama-65m-series-b-9m-users)

• US-China open-model benchmark rivalry intensified in 2026 (Nemotron 3 Ultra vs Kimi K2.6)

• Enterprise adoption via Hugging Face, OpenRouter, Ollama, and NIM microservices

• Benchmark predicts open-weight models will generate supermajority of tokens within 18–24 months

• Complements but does not replace proprietary frontier APIs for most teams

Related

• open-secure-ai-defense

• hugging-face-ai-agent-security-incident

• glm-5-2

• hugging-face

• moonshot-ai

• inkling

• thinking-machines-lab

• us-open-weights-race-2026

• ollama

• open-source-llm

• open-source-ai

• on-device-ai

• llm-infrastructure

• nemotron-3-ultra

• nemotron-3-embed

• qwen-38

• mellum2

• us-china-ai-competition

Sources

• 2026-07-17-moonshot-kimi-k3-28t-open-source

• 2026-07-09-ollama-65m-series-b-9m-users

• 2026-06-01-nvidia-nemotron-3-ultra-computex

• 2026-06-02-jetbrains-mellum2-open-source

• 2026-07-23-cisco-antares-open-weight-vuln-localization

• 2026-08-03-alibaba-qwen3-8-max-launch