Skip to content

AI · 22 November 2026 · 4 min read

November AI Model Releases Hit 45 Across 34 Vendors as Open-Weight Sparse MoE Models Target Consumer GPUs

More than half of November’s 45 model launches carry open weights, led by sparse architectures like Edge0-145B that run 145B total parameters on a single 16 GiB consumer graphics card.

A trade-show booth labelled NVIDIA Vera Rubin NVL72, with an open liquid-cooled rack and visitors in front.
A photograph of an NVIDIA Vera Rubin NVL72 rack on an ASRock display. It is not one of the 45 model releases, and it is not Edge0-145B.

Lede & Nut Graf

Artificial intelligence (AI) model release velocity reached 45 launches across 34 vendors through November 22, 2026—with 23 of those releases shipping with open weights (AIFOD Realtime Tracker, Nov. 22, 2026). Even as cloud hyperscalers roll out flagship reasoning models, the defining engineering trend of the month is aggressive parameter sparsity that pushes 100-billion-plus-parameter intelligence onto single-card local hardware.

For enterprise chief information officers (CIOs) and cloud infrastructure investors, the rapid maturation of open-weight Mixture-of-Experts (MoE) models directly challenges the pricing power of frontier application programming interfaces (APIs).

Technical & Financial Breakdown

At the top of the proprietary cloud stack, pricing remains heavily segmented by reasoning depth. OpenAI’s GPT-6.2 Astra carries list rates of $8.00 per million input tokens and $40.00 per million output tokens, posting a score of 62 on the Artificial Analysis Intelligence Index, while Google DeepMind’s Gemini 4 Neon prices at $3.00/$15.00 per million tokens and scores 61 at xhigh reasoning (AIFOD Realtime Tracker, Nov. 22, 2026).

In parallel, open-weight developers are attacking inference memory bandwidth—the primary bottleneck of local deployment—by activating only a small fraction of total weights per token. Exemplifying this shift, Edge0-145B packs 145 billion total parameters but activates just 10.12 billion parameters per token, allowing the model to run locally on a single consumer graphics processing unit (GPU) with 16 gibibytes (GiB) of video random-access memory (VRAM) and 20 GiB of system RAM.

Second-Order Implications

When a single 16 GiB workstation card can host a 145B-parameter MoE model with zero per-token API fees, privacy-sensitive enterprises in healthcare, defense, and finance gain a compelling off-ramp from cloud inference bills. Combined with vertical startups like Model ML shifting workloads to fine-tuned open-weight models at a fraction of frontier pricing, proprietary labs face intensifying margin compression on routine reasoning and coding traffic.

Caveats & Open Questions

Offloading inactive MoE expert weights between 16 GiB of VRAM and 20 GiB of system RAM across PCIe lanes typically incurs steep decode-latency penalties compared with pure high-bandwidth memory (HBM) execution. Independent tokens-per-second throughput benchmarks and head-to-head reasoning scores for Edge0-145B against Gemini 4 Neon and GPT-6.2 Astra were not listed in the tracker summary.