Total 75 · Clear filters
On 26 October 2026 PFN released the weights of PLaMo 3 Translate, a 53-language translation model built on a base model from joint research with NICT, in four sizes from 0.8B to 31B and several quantised formats.
On 15 October 2026, IFM and Inception42 announced Jais 3, six Arabic-centric language models from 900 million to 375 billion parameters plus a speech recogniser, evaluated on 24 audited Arabic benchmarks.
On 3 September 2026 the Institute of Foundation Models announced K2 Horizon, a fleet of six models released together, each shipped with its training data, recipe and evaluations, and steered by a router that picks the cheapest adequate model per task.
In September 2026 StepFun released Step 5 Preview as its new flagship foundation model. It is a sparse MoE with 600 billion total parameters, supports up to 1 million tokens of context, and is offered through the StepFun Open Platform API.
Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter model, as a cloud service on 3 August 2026, and published open weights for a variant of it on 12 August 2026.
Baidu released ERNIE 5.1 on 9 May 2026, a model derived from ERNIE 5.0's pre-training with about a third of its total parameters and much lower training cost, after a preview build topped Chinese models on LMArena.
Shengshu Technology announced a Series B round of nearly 2 billion yuan led by Alibaba Cloud, money it says will go to foundation world models for both digital generation and physical action.
On 9 April 2026 LG AI Research released EXAONE 4.5, its first open-weight multimodal model: a 33B vision-language system aimed at document understanding, with a 256K context window and the earlier EXAONE licence.
Skild AI, a maker of general-purpose robotic software, raised $1.4 billion in a Series C led by SoftBank, valuing the company at more than $14 billion.
Boston Dynamics and Google DeepMind announced a partnership at CES 2026 to apply DeepMind's AI foundation models to Boston Dynamics robots, with the humanoid Atlas as the first test case.
Upstage published Solar Open 100B on 10 December 2025: a 102.6-billion-parameter model that activates 12 billion parameters per token, trained on 19.7 trillion tokens and released as open weights under a company licence.
On 19 November 2025 Meta released SAM 3, which detects, segments and tracks objects from text or visual prompts, and SAM 3D, which reconstructs 3D objects and human bodies from a single image. Both were released as open models.
On 13 November 2025 Baidu unveiled ERNIE 5.0 at Baidu World 2025, describing it as a natively omni-modal model. It ranked eighth on the LMArena text leaderboard after leaving preview.
On 28 October 2025 NVIDIA released Isaac GR00T N1.6, an open model for humanoid robots that adds a reasoning vision-language model, Cosmos Reason, so vague instructions can be turned into step-by-step action plans.
In September 2025 Alibaba released Qwen3-Max, a proprietary model with over 1 trillion parameters trained on about 36 trillion tokens. It became the flagship of Alibaba's closed Max tier, released alongside several open Qwen3 models.
On 28 August 2025 Microsoft AI announced its first in-house models, the MAI-Voice-1 speech model and the MAI-1-preview language model. Coverage read the release as a first step to reduce dependence on OpenAI.
Infinigence AI released Megrez2-Preview on 22 July 2025 and the finished Megrez2-3x7B-A3B in September, an on-device mixture-of-experts model in which three adjacent layers share one pool of 64 experts.
On 21 July 2025 NAVER published HyperCLOVA X SEED 14B Think, a reasoning model built on top of a pruned and distilled model, with training reported at 68,049 A100-80GB GPU hours.
LG AI Research published EXAONE 4.0 on 15 July 2025, a two-size family with a switchable reasoning mode. The 1.2B model targets on-device use, and a licence change allows commercial deployment through partners.
On 23 May 2025 Sarvam AI released Sarvam-M, an open-weights Indic model built on Mistral Small, publishing the weights, a playground and APIs at the same time.
On 1 May 2025, TII released Falcon-H1, the first Falcon family to combine attention layers with state space models in parallel, shipping six sizes from 0.5B to 34B parameters at once.
