Open Weights Outside China: Llama, Mistral, Gemma — 2026-10-03

Top developments

Upstage releases Solar Mini 4: MoE model optimized for on-premises deployment

Upstage unveiled Solar Mini 4 on October 1–2, a 35-billion-parameter model with a Mixture-of-Experts (MoE) structure that activates only 3 billion parameters during inference, reducing computational load for enterprises. The model supports quantization to run on a single GPU, positioning it for on-premises adoption in corporate environments. Open weights and model code were released alongside benchmarks.

Mistral AI signals major model announcement to compete with US frontier models

Mistral announced on September 30 that a new model is in development and will "significantly close the gap" with American models, positioning itself as France's answer to OpenAI and Google. The statement came alongside news of a Morocco partnership releasing two open-source models for Darija (Moroccan Arabic dialect) identification and transcription, with publication on open-source platforms confirmed.

Cloudflare launches open-weight decision models, claiming benchmarks above proprietary systems

On October 1, Cloudflare released Clef and Clef-flash, two open-weight decision-making models designed for constrained text generation. The larger model already claims benchmark superiority over proprietary reasoning systems in early evaluations. The models are released under permissive open-weight licenses.

Qwen dominates local inference market: 39.6M monthly GGUF downloads vs. Llama's 7.5M

Hugging Face data shows Qwen leads GGUF quantized downloads at 39.6 million per month—nearly twice Gemma's 20.8 million and more than five times Llama's 7.5 million. The dominance reflects both superior inference optimization and adoption momentum across local deployment platforms. Despite Llama-derived repositories outnumbering Qwen's, the performance gap persists.

Local view

Korean media: IT Chosun and Money Today emphasized Upstage's efficiency claims, highlighting that Solar Mini 4 can run AI tasks on a single GPU with quantization—a breakthrough for on-premises corporate AI in Korea's enterprise sector. Chosun Biz framed it as Korean LLM competitiveness against NVIDIA and Alibaba models.

French media: Journal du Geek positioned Mistral's announcement as a strategic European response to US AI dominance, citing the new model as critical for France's AI sovereignty narrative. King of Geek and Agence Ecofin covered the Morocco partnership as evidence of Mistral's regional expansion strategy beyond Europe.

Context & numbers

GGUF market concentration: Qwen's 39.6M monthly downloads represent consolidation around a single ecosystem leader for quantized inference.

Open-weight licensing trends: Qwen3 carries Apache 2.0 (permissive commercial use); Gemma 4 maintains Google's proprietary-friendly terms; Mistral models use Apache 2.0. License clarity and commercial flexibility drive enterprise adoption.

Model size trends (September 2026): MoE structures dominate efficiency announcements—Xiaomi MiMo-V2.6-Flash (309B total / 15B active) and Solar Mini 4 both prioritize reduced active parameters for resource-constrained deployment.

On the radar

• NVIDIA Nemotron ecosystem: NVIDIA's open Nemotron datasets and models (announced via Hugging Face September 28) position the vendor as an alternative to pure open-weight labs, with permissive data licensing and evaluation frameworks shipping alongside weights. Watch for enterprise adoption curves over Q4 2026.

• Cloudflare sovereign AI positioning: Cloudflare's October 1 statement on sovereign AI (alongside Clef release) signals intent to decouple regional AI compliance from US policy, echoing European and Asian regulatory concerns. Model releases may accelerate if geopolitical licensing disputes intensify.

• Mistral Large 3 launch timing: CEO statements suggest the next flagship model is imminent but unscheduled as of October 3. Watch official announcements and Hugging Face model card publication for release date and benchmark details.