Atlas / Open-weight and challenger labs

Open-weight and challenger labs

AI Market Atlas · by Pro Pluribus

Models you can download and run yourself, plus well-funded challengers. China now holds most of the top open-weight spots.

20 companies and products

Updated 16 Oct 2026. Select one for its offerings, news and links.

Open-weight and challenger labs

China
DeepSeekUpdatedV4 and V4.1, the price-setter for open frontier models. Raising $12B+ pre-IPO led by Tencent and CATL, with a listing targeted for early 2027.QwenUpdatedQwen3.8 is current; Qwen 4 was previewed in four tiers (22 Sep) but not yet released. The most-used base model for fine-tuning.KimiUpdatedKimi K3, the largest open-weight model. Moonshot is valued at ~$50B and targets a Hong Kong IPO in Q1 2027.GLMUpdatedGLM-5.3: billed as the strongest open-weight coding model.MiniMaxM3 open model, plus Hailuo video and consumer apps.HunyuanHy4 preview, the cheapest flagship-class open model.LongCatTopped OpenRouter usage anonymously before it was even announced.Doubao · SeedChina's #1 consumer chatbot. Flagship models stay closed.Xiaomi MiMoA phone maker now holds the top open-weight score: MiMo-V2.6-Pro leads the Artificial Analysis open-weight rankings.
US and Europe
Mistral AIUpdatedEurope's sovereign-AI champion. Raised EUR 3B at EUR 21B (Sep); previewed Large 4, a 1T-parameter open model (6 Oct).Gemma 4The strongest open small-model family from a US lab.Nemotron 3UpdatedThe largest US builder of open-weight models, with open data and training recipes.gpt-ossOpenAI's only open-weight models, now a year old.Thinking Machines LabInkling open models plus Tinker, a service for fine-tuning your own model.CohereUpdatedEnterprise and sovereign specialist. Merging with Germany's Aleph Alpha at a $20B valuation (16 Sep); shipped North 2.OLMoFully open: weights, training data and code.Reflection AIUpdatedShipped its first model on 16 Oct: Beam, a 501B open-weight model. NVIDIA is reportedly in talks to acquire it.Arcee AIUpdatedUS open-weight challenger: $150M+ Series B at $1B pre-money (Sep); says it trained four open models for $20M.PoolsideCoding-focused model lab. NVIDIA is paying $6B to license its training stack and investing $1B at a $12B valuation.JevUpdatedA new kind of model that returns typed decisions, not text. $870M Series A at $7.5B (9 Oct), under a month after launch; open rivals already appearing.

Compiled from public sources by Pro Pluribus; updated 16 Oct 2026. General information, not advice. Company names and logos belong to their owners and are used to identify them, without affiliation or endorsement. Terms