China Merchants
Nov 4, 2026

Lion Rock Lab unveils embodied reasoning VLM with roughly 3B activated parameters

Lion Rock Artificial Intelligence Laboratory, which operates under China Merchants Advanced Technology Research Institute, has released an embodied reasoning vision-language model called Lion-ER-35B-A3B under the Chinese name "Yuwu." The model uses a mixture-of-experts, or MoE, architecture, with about 35 billion total parameters while activating only around 3 billion parameters for each token. According to the release cited by Techub News, the training set was built through a capability-driven embodied data engine and spans eight top-level domains, including objects, space, and temporal reasoning, along with 39 secondary capabilities. The supervised dataset contains 312,000 samples. On 10 public embodied benchmarks, the model posted gains across all reported metrics, with an equal-weight composite score of 65.29. On Where2Place, it scored 64.28, up 9.30 points from its base model. The item cited Quantum Bit as the source of the underlying information.

370
Lion Rock Lab unveils embodied reasoning VLM with roughly 3B activated parameters
NVIDIA
Nov 3, 2026

NVIDIA details three inference optimizations for Ling, lifting throughput 42.5% on DGX Spark

NVIDIA said in a technical blog post that it tested three optimizations for Mixture-of-Experts inference on DGX Spark using Ling-3.0-flash, a model from Ant Group-backed inclusionAI. The model is described as a 124B-parameter MoE system with 5.1B parameters activated per token. The optimizations listed were the Humming MoE kernel, an FP8 LM head, and MTP speculative decoding. According to the figures cited by MarkTechPost, results from a 1,319-question GSM8K test showed decoding throughput rising from 37.81 tokens per second to 53.89 tokens per second, a 42.5% increase. Average latency fell 24.5%, while accuracy stayed nearly unchanged at 96.89% and 96.82%, respectively. The post focused on inference efficiency gains rather than changes to model quality, with the reported accuracy figures showing little difference between the compared setups.

780
NVIDIA details three inference optimizations for Ling, lifting throughput 42.5% on DGX Spark
Prime Intellect
Oct 29, 2026

Prime Intellect launches Prime Mega MoE, reporting a 3.11x gain in MoE training forward speed

Prime Intellect, an AI infrastructure company, has introduced Prime Mega MoE, a trainable extension built on DeepSeek’s open-source Mega MoE kernel. The release adds fast, overlapped backward propagation to a kernel that previously supported only forward propagation, allowing it to be used for training rather than inference alone. According to Prime Intellect, the full training step for the Qwen3 MoE model runs up to 34% faster inside its prime-rl trainer, and the gain reaches as high as 55% when compared with TorchTitan. The company also said the kernel supports NVIDIA Hopper and Blackwell GPUs, along with BF16, MXFP8, and blockwise FP8 data types. Prime Intellect described Prime Mega MoE as a fused and overlapped MoE training kernel designed to address GPU idle time caused by communication bottlenecks during MoE training. It said the result is a direct reduction in required GPU hours for teams training mixture-of-experts models. The item was cited by MarkTechPost and published by Techub.

650
Prime Intellect launches Prime Mega MoE, reporting a 3.11x gain in MoE training forward speed
Ant Group
Oct 28, 2026

Ant Group’s Bailian open-sources Ling-3.1-flash-1019 for Personal Agent use cases

Ant Group’s Bailian large model team has open-sourced Ling-3.1-flash-1019, a new version aimed at Personal Agent scenarios. The release builds on Ling-3.1-flash and scored 44 points on the AAII v4.3.2 intelligence index, up 3 points from the previous version. Ant said the model is designed to improve performance on complex workflows and multi-step tasks. Alongside it, the company also open-sourced the base model Ling-3.1-flash-base. That model uses a hybrid linear Mixture-of-Experts, or MoE, architecture and has roughly 560B parameters in total. According to figures shared by Bailian, Ling-3.1-flash-base outperformed DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base in 32 of 47 benchmark tests spanning eight domains. Bailian said it hopes the open-source release will help developers explore more application possibilities. The item was cited by Techub News and attributed to QbitAI.

510
Ant Group’s Bailian open-sources Ling-3.1-flash-1019 for Personal Agent use cases
Zhipu
Oct 23, 2026

Zhipu’s compensation notice hints at a new model as GLM-5.5 speculation builds

A compensation notice sent to some GLM Coding Plan users has fueled fresh speculation that Zhipu may be preparing to release a new model, widely discussed online as GLM-5.5. The notice addressed a reported “capability downgrade” issue affecting some users and said that after the new model launches, users will receive an additional quota reset opportunity. While the company has not disclosed the next model’s name or launch date, that wording has drawn attention because it is one of the few recent references to an upcoming release. Discussion around the next GLM model has centered on three areas: parameter scale, architecture changes, and whether Loop Transformer could be introduced. Online rumors cited in the source suggest the next model may reach 2.5 trillion to 3.5 trillion parameters, compared with roughly 743 billion parameters for GLM-5.3, which Zhipu released in August using a mixture-of-experts, or MoE, architecture. The source also points to GLM-5.3-Flash, which introduced a hybrid design combining linear attention and sparse attention, along with mHC technology, and reportedly delivered stronger benchmark results than GLM-5.2 at about one-tenth of the inference price. Zhipu has not officially confirmed any GLM-5.5 specifications.

1010
Zhipu’s compensation notice hints at a new model as GLM-5.5 speculation builds
Recursive
Oct 22, 2026

Recursive unveils Flint, a 40B-parameter MoE model built for local coding

AI startup Recursive has introduced Flint, its first language model, according to Techub News, citing The Decoder. The model uses a 40 billion-parameter mixture-of-experts, or MoE, architecture and is designed specifically for local coding workloads. The report says Flint was developed by former DeepMind researchers. Recursive also said Flint was built with help from its in-house AI system, Eureka Machine. Development took four months, and the company said the system can independently test and optimize model architectures. The release marks the debut of Recursive’s first model and outlines the company’s focus on coding use cases that run locally rather than in a broader general-purpose setting.

1040
Recursive unveils Flint, a 40B-parameter MoE model built for local coding
MemT 1.0
Oct 20, 2026

MemT 1.0 open-sourced with 36B and 288B models, reaching near-300B memory-native validation

MemT 1.0, a memory-native model series, has been released as open source with two versions: the 36B-parameter Nano model and the 288B-parameter Flash model. According to the release cited by Techub News, the series adds a parallel memory-compute path inside Transformer modules, allowing experience to be encoded as internal dynamic states and read or written through training. The team said this marks the first engineering validation of a memory-native design at nearly 300B parameters. On benchmarks including SWE-bench Verified, MemT-1.0-Nano posted a 69.73% average success rate on key breakpoints. The share of tasks that succeeded in all three runs rose from 54.00% under a no-experience baseline to 62.80%. In tasks such as maze exploration, the model was able to use experience from the previous round to improve later actions. MemT 1.0 uses Gated DeltaNet-2 for per-token memory updates and introduces a Mixture-of-Storage approach built on a Mixture of Experts, or MoE, architecture to expand memory capacity. The team trained the memory system on 34 billion tokens across continued pretraining and post-training stages.

780
MemT 1.0 open-sourced with 36B and 288B models, reaching near-300B memory-native validation
Microsoft
Oct 16, 2026

Microsoft says it has finished pretraining MAI-Base-2, a 2.8 trillion-parameter model

Microsoft said it has completed pretraining for MAI-Base-2, its latest in-house foundation model, with total parameters reaching 2.8 trillion. The model uses a mixture-of-experts, or MoE, architecture and activates 80 billion parameters per inference step. According to the company, training ran for 38 days across August and September on 16,000 NVIDIA GB200 GPUs and processed about 50 trillion tokens. Microsoft also said the model introduces native vision pretraining for the first time, with visual data accounting for 18% of the training mix. In internal testing disclosed by Microsoft, MAI-Base-2 outperformed models including DeepSeek-V4-Pro-Base, Kimi K3 and Qwen3.8 on text prediction benchmarks covering coding, question answering, mathematics and science and engineering subjects. Microsoft added that training a model at the same level as DeepSeek-V4-Pro-Base would require only about one-quarter to one-twelfth of the compute under its method, though it kept scaling MAI-Base-2 further. Mustafa Suleyman, Microsoft’s AI chief, said at an internal meeting on Oct. 15 that the pretrained base surpasses the OpenAI model behind GPT-5.6 Sol while using one-quarter of the training compute. The model remains in post-training, with deployment planned first in Microsoft products and later through Foundry to enterprise partners.

1160
Microsoft says it has finished pretraining MAI-Base-2, a 2.8 trillion-parameter model
Total 41 items
Go topage