Open Weights Outside China: Llama, Mistral, Gemma — 2026-10-18
Meta released its massive 816B-parameter "Muse Spark Open" model, optimized for Apple Silicon and claiming top performance among US open-weight models. NVIDIA simultaneously launched the Nemotron 3.5 series, including a 120B vision-language model and smaller edge variants, while Mistral Large 4 open weights are anticipated by late October with an unknown license. The open-weight ecosystem continues to grow, with Qwen maintaining significant download dominance and new benchmarks highlighting the performance of models like GLM-5.3 and Kimi K3.
Open Weights Outside China: Llama, Mistral, Gemma — 2026-10-18
Top developments
Meta Launches 816B "Muse Spark Open" Optimized for Apple Silicon
Meta released the weights for "Muse Spark Open" on October 15, 2026, a native multimodal Mixture-of-Experts (MoE) model with 816 billion total parameters and 42 billion active parameters. The model features a 1 million token context window and is specifically optimized to run locally on Apple Silicon, with Meta demonstrating it on two M5 Ultra Mac Studios. The Guardian notes Meta’s claim that it excels in "multi-step reasoning" and "complex software engineering," positioning it as the highest-performing US-made open-weight model, though it uses a custom license with conditions for commercial use.
NVIDIA Releases Nemotron 3.5 Super VL and Bolt Edge Models
On October 15, 2026, NVIDIA unveiled three new open Nemotron models: the 120B-parameter Nemotron 3.5 Super VL, and two smaller "Bolt" models for edge and local hardware. The Super VL is a multimodal model designed for specialized, always-on agents, while the Bolt variants target data-center-to-desktop scenarios. ZDNET highlights that these releases are part of a broader wave of US open-weight initiatives, aiming to provide alternatives to Chinese models.
Open-Weight Models Capture Significant Market Share
A recent report from The Register indicates that open-weight models now handle 71.5% of Vercel gateway tokens and account for nearly a quarter of estimated spending in their platform. This statistic underscores the growing economic viability and developer adoption of open-weight models like Llama, Mistral, and Qwen, moving beyond experimental use to production-critical workloads.
Mistral Large 4 Open Weights Expected Late October, License Unknown
While Mistral Large 4 was announced earlier in October, the open weights are expected to be released by late October 2026. French tech outlet Stephane Larue reports that the specific license governing the weights remains unknown, creating uncertainty for enterprise adoption. The model is trained on 4,000 Nvidia GPUs in Europe and is positioned to lead open-weight models outside China.
Reflection AI Ships Beam 1.0 with Apache 2.0 License
Reflection AI released Beam 1.0 on October 16, 2026, under the permissive Apache 2.0 license. Unlike some competitors, Beam 1.0 has no set API price, focusing entirely on open-weight distribution. This release contrasts with Qwen3.8-Max, which has established API pricing, highlighting different monetization strategies within the open-weight sector.
Local view
Japan: ITmedia and Unite.AI Highlight Meta's Mac Studio Optimization
Japanese tech outlets ITmedia and Unite.AI focused heavily on the hardware implications of Meta's Muse Spark Open. ITmedia emphasized the claim that the model runs on two M5 Ultra Mac Studios, framing it as the "highest performance US-made open model." Unite.AI noted the technical specs (816B total/42B active params) and the native multimodal capabilities, while also pointing out the custom license conditions for commercial use.
France: Le Monde and Stephane Larue Focus on Sovereignty and Licensing
Le Monde framed Mistral Large 4 as a strategic move to "reduce the gap" with Chinese competitors, emphasizing the European training infrastructure. Stephane Larue, a prominent French AI analyst, focused on the business implications, noting that the unknown license for the open weights is a critical missing piece for French enterprises seeking sovereign AI solutions.
Korea: Rabbithup Releases GLM-5.2 Draft Model for Korean Language
Korean outlet NewsPim reported that Rabbithup released "GLM-5.2-NVFP4-DFlash-ko" on Hugging Face. This is a draft model designed to accelerate inference for Korean language queries and agent tasks, based on Z.ai's GLM architecture. This signals continued localization efforts for open-weight models in the Korean market.
Context & numbers
- Download Dominance: Qwen continues to lead in local inference downloads, with 39.6 million GGUF downloads per month, nearly double Gemma's 20.8 million and over five times Llama's 7.5 million (data from August 2026, still relevant context).
- Benchmark Scores: In October 2026 comparisons, GLM-5.3 and Kimi K3 lead the Artificial Analysis Index with scores of 45 and 44 respectively. Meta's Muse Spark Open scored 39 but took first place in tool-use tests.
- Market Sentiment: A Manifold market prediction shows 86% odds on open-source AI being a significant data-center play, driven by releases from Reflection AI (Beam) and Mistral (Large 4).
On the radar
- Mistral Large 4 Weights Release: Keep an eye on the end of October for the actual weight drop and license text for Mistral Large 4.
- Moondream 3 Endpoint Shutdown: The hosted endpoint for Moondream 3 Preview will shut down on November 13, 2026, requiring users to migrate to local alternatives or other providers.
- Jais 3 vLLM Support: Jais-3-375B-A23B weights were released in late September, but vLLM support recipes were dated October 5, with press releases following on October 15, indicating a lag in major serving framework support for new open models.
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
