Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-27
A quiet week for new Western open-weight launches, but rich in strategy and policy discussion. Nathan Lambert of Interconnects published congressional-testimony-grade analysis of the open-model balance of power, while Mistral CEO Arthur Mensch faced pointed questions in Le Monde about falling behind US and Chinese rivals. In Korea, Naver Cloud's HyperCLOVA X and LG AI Research's Exaone secured a government cyber-security AI project, and Japan's domestic LLM landscape saw its first major leaderboard reshuffle of the fall.
Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-27
Top developments
"The current balance of power in open models" published as Congressional testimony
Nathan Lambert (Interconnects) released, about six days ago, an expanded version of testimony he prepared for Congress analyzing where open-weight AI stands globally. The piece frames the strategic debate over which jurisdictions — the US, Europe, East Asia — can sustain frontier-grade open-weight ecosystems featuring models like Llama, Gemma, Mistral and Nemotron. It matters because open weights are now explicitly a policy question in Washington, not just a release-cycle story — the first time this signal has had a Congress-facing primary document to report on.
Arthur Mensch defends Mistral's open-weight strategy in Le Monde interview
In a Le Monde interview published September 24 (three days ago), Mistral AI's CEO argued "AI is software. It can be controlled," pushed back on what he called manipulated risk discourse by American giants, and defended his strategy as critics accuse the company of falling behind US and Chinese rivals. A Medium analysis (The AI Brief) adds that after DeepSeek's rise, Mistral — once the lab with the top open-source model — is now beaten by Gemma 4 on the open side of its lineup. Mistral's open family still includes Apache-licensed Small, Nemo and Ministral models alongside its proprietary Large flagship.
Korea: Naver Cloud and LG AI Research land cyber-security AI foundation model project
Korean outlets reported on September 22 that the Ministry of Science and ICT held a briefing for the "cyber-security-specialised AI model development project," with Naver Cloud's HyperCLOVA X and LG AI Research's Exaone competing — Money Today's piece circulated via Supple frames it as "Naver defends, LG attacks." TechM reports the two will build 10 months of foundation-model development targeting attack and defence specialisation, with LG's attacking Exaone aiming to match GLM-family level on vulnerability-discovery benchmarks. This is a significant government commitment to sovereign non-Chinese open AI.
Japan: domestic LLM rankings reshuffled; government picks seven models
Eques published a fresh comparison (within the last day) of seven government-selected domestic Japanese LLMs, aimed at enterprises that cannot use foreign cloud AI — reflecting strong demand for sovereign Japanese weights. Qualiteg's Nejumi Leaderboard 4 analysis noted the Japanese-language LLM ranking saw its No.1 change within two months, underscoring how fast the domestic scene is moving. NTT Data's tsuzumi 2 remains the Digital Agency's Government AI trial model.
Local view
- Korea (Money Today via Supple, September 22): coverage centres on the "K-myth" framing of Naver attacking and LG defending in the state-funded cyber-security AI race, presenting HyperCLOVA X and Exaone as Korea's answer to foreign frontier models.
- France (Le Monde, September 24): Mensch's interview is a defensive repositioning of Mistral amid claims it lags US and Chinese rivals; French tech commentary (duperrin.com, September 24) analyses Mistral's strategy as "models, compute, integration and reversibility" for European enterprise AI.
- Japan (Eques, September 27): enterprise-facing guidance warns firms that "we can't send data abroad" is the typical reason procurement stalls, driving interest in government-vetted domestic open models.
Context & numbers
- Monthly GGUF downloads (Hugging Face, Summer 2026 baseline for comparison): Qwen ~39.6M/month, Gemma 20.8M, Llama 7.5M — Llama-derived repos slightly outnumber Qwen's, so the gap is not supply-side.
- Mistral's open-weight family retains Apache 2.0 licences on the official Hugging Face model cards; Gemma 4 26B A4B is also listed with Apache 2.0 and described as a practical local-model pick.
- Qwen derivatives grew at roughly 180–210 new repositories per day through the first seven months of 2026, showing sustained fine-tune ecosystem momentum.
- Community quant ecosystem in the past week: unsloth/Qwen3.8-27B-GGUF passed 6.9M downloads; ternary 2-bit, GSQ and mixed-precision quantisation are becoming standard deployment tools.
On the radar
- Watch for official Korean government announcements following the September 22 cyber-security AI project briefing — 10-month foundation-model timelines could produce open deliverables from Exaone and HyperCLOVA X.
- Japanese-language LLM leaderboard churn: after a No.1 change in the September 1 Nejumi edition, a October revision could reshuffle the domestic open-model ranking again.
- Rumour (flagged as such): Medium's The AI Brief suggests pressure is mounting on Mistral to answer Gemma 4's open-weight benchmark lead with a stronger open release; no official announcement exists yet.
- No new open-weight releases from Meta, Google, Nvidia, or TII were verifiable within the past 7 days in these results; expect the next Llama/Gemma/Falcon news cycle to be tracked against Lambert's Congressional testimony framing.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
