Liquid AI
Oct 14, 2026

Liquid AI updates 24B local model, posts higher Agent scores than larger Qwen model

Liquid AI has introduced LFM2.5-24B-A2B, a new local model aimed at improving on-device Agent reasoning and tool use. The model keeps the same 24 billion total parameters as its predecessor while activating only about 2 billion parameters per run. Compared with LFM2-24B-A2B, the new release expands pretraining data from 17 trillion to 33 trillion tokens and increases context length from 32K to 131K. Liquid AI said the model is trained to reason before answering and was reinforced directly inside Agent frameworks including OpenClaw, Hermes Agent, Pi, and OpenCode to practice multi-step tool calling and result handling. In the company’s demo, Hermes Agent ran on a MacBook without using a cloud model API, searched trending Hugging Face projects, compiled a briefing, and continued after follow-up prompts. Liquid AI’s tests said the model generated 184 tokens per second on Apple’s M5 Max, with peak speed reaching 1.7x that of comparable MoE models. On Agent benchmarks, it scored 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, and 27.69 on AutomationBench, more than three times Qwen’s result. On SWE-bench Verified, however, it scored 51%, below Qwen’s 66.2%.

1900
Liquid AI updates 24B local model, posts higher Agent scores than larger Qwen model
Reflection
Oct 6, 2026

Reflection unveils 501B-parameter Beam, trailing the latest top Chinese open-source models

Nvidia-backed AI startup Reflection has introduced Beam, its first open-weight model. Built on a mixture-of-experts, or MoE, architecture, Beam has 501 billion total parameters and activates 23 billion parameters per inference step. The model is aimed at coding, reasoning and agent-based workloads. Based on metrics released by Reflection, Beam performs close to GLM-5.2 overall, but it does not yet match the newest leading open-source models from China. On DeepSWE, Beam scored 44.4, compared with 44.0 for GLM-5.2 and 51.0 for Qwen 3.8-Max. GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash posted 61.0, 68.0 and 74.2, respectively. Reflection also disclosed the scale of Beam’s training. Pretraining used 6,144 Nvidia GB300 GPUs and processed 23.8 trillion tokens in under four weeks. That was followed by four continuous weeks of reinforcement learning on 10,500 GB300 GPUs, producing more than 100 million rollouts. The company said this ranks among the largest reinforcement learning training runs in the public record, to its knowledge. Beam remains in final safety testing, with only a small number of users getting early access. Reflection said it plans to release the full weights, a technical report and a model card later this month under the Apache 2.0 license.

1120
Reflection unveils 501B-parameter Beam, trailing the latest top Chinese open-source models
Reflection AI
Oct 6, 2026

Reflection AI unveils open-weight model Beam with a focus on coding and agent workloads

Reflection AI, a U.S. startup backed by Nvidia, said on Oct. 5 that it has introduced Beam, its first open-weight model, and plans to release the weights later this month under an Apache 2.0 license alongside a technical report and model card. Beam uses a sparse mixture-of-experts architecture with 501 billion total parameters and 23 billion active parameters per inference step, according to the company. Reflection said Beam is aimed at software development, reasoning, and AI agent tasks. The startup claimed the model posts advanced reasoning benchmark results comparable to Z.ai’s open model GLM-5.2 while using 3x to 4x less inference compute. It also said Beam is competitive with larger open models in coding and agent tasks and comes close to Alibaba’s Qwen 3.8-Max, though it acknowledged that Moonshot AI’s Kimi K3 still leads on raw capability. The company disclosed benchmark scores including 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1. TechCrunch noted that the results have not been independently verified. Reflection also detailed the training run, saying Beam was pretrained on 23.8 trillion tokens and later trained with reinforcement learning on 10,500 Nvidia GB300 GPUs over four weeks. The startup, founded in 2024 by two former Google DeepMind researchers, has raised about $4.7 billion, according to PitchBook data cited by TechCrunch.

890
Reflection AI unveils open-weight model Beam with a focus on coding and agent workloads
Aleph Alpha
Oct 4, 2026

Aleph Alpha open-sources Kolibri with 3.46B active parameters and up to 1M context

German AI company Aleph Alpha has released and open-sourced Kolibri, a bilingual English-German model built on a mixture-of-experts architecture. The model has 78.1 billion total parameters, while activating about 3.46 billion parameters per token. Its weights have been uploaded to Hugging Face under the Apache 2.0 license. Kolibri supports tool calling and adjustable reasoning modes. Its context window can scale to roughly 1 million tokens, though the company said daily deployment should stay within 262,000 tokens. Aleph Alpha also said the longest context length used in actual training was 262,000 tokens. In the company’s official evaluation, Kolibri scored 96.9% on AIME 2025 and 85.9% on LiveCodeBench v6. On combined English and German results, it slightly outperformed Qwen3.5 35B-A3B, which has a similar active-parameter scale of about 3B, but remained behind Qwen3.8, a 27B dense model.

1190
Aleph Alpha open-sources Kolibri with 3.46B active parameters and up to 1M context
DeepSeek
Oct 1, 2026

DeepSeek open-sources Ascend infrastructure components with APIs aligned to Nvidia versions

DeepSeek said it open-sourced a batch of infrastructure components for Huawei’s Ascend platform on Sept. 30, publishing the code on GitHub and describing the new releases as one-to-one counterparts to components it had previously released for Nvidia GPU systems. The package includes DeepGEMM Ascend, a matrix multiplication library whose documentation says the first version supports the Ascend 950 series and remains fully compatible with the original DeepGEMM API, allowing developers to keep the same workflow across platforms. It also includes DeepEP-Ascend, a communication library for training and inference that provides expert-parallel communication operations for mixture-of-experts models and exposes a buffer API aligned with the Nvidia version of DeepEP. According to QbitAI, the open-source release also covers TileLang, FlashMLA, TileKernel and DeepSelect. Documentation for DeepEP-Ascend lists benchmark results from tests on Ascend 950DT with CANN 9.2.0, showing dispatch bandwidth of 373-375 GB/s and combine bandwidth of 345-347 GB/s at eight expert-parallel units, with lower figures at 128 units. The documents also note that full bandwidth depends on a Huawei commercial firmware release for Atlas 850E expected around mid-October 2026.

1380
DeepSeek open-sources Ascend infrastructure components with APIs aligned to Nvidia versions
NaiveAI
Sep 28, 2026

NaiveAI open-sources first model as AI helps drive 151 optimization runs in six days

NaiveAI, founded by Tsinghua University associate professor Dai Jifeng, has released Naive-N0.5-Flash and opened its model weights under the MIT license. The model is based on Xiaomi’s MiMo-V2.5 Base and keeps the original mixture-of-experts, or MoE, architecture while changing the attention design. According to the team, global attention was replaced with sliding-window attention and DeepSeek sparse attention, followed by continued training on 3.25 trillion tokens. The updated model natively supports a 1 million-token context window, which the company said cuts compute costs for very long text processing. NaiveAI also said AI systems were directly involved in development, handling coding, experiments, result analysis and later optimization, while human researchers set direction and made key decisions. Using its NaiveRT inference system as an example, the team said it completed 151 optimization rounds over six days, with 63 adopted. The company also disclosed a peak inference speed of 2,122 tok/s under a specific setup, while saying standard mode runs at about 50 tok/s per user.

1200
NaiveAI open-sources first model as AI helps drive 151 optimization runs in six days
Vitalik Buterin
Sep 26, 2026

Vitalik says offline mobile AI apps have improved, but still trail laptop models

Ethereum co-founder Vitalik Buterin said mobile offline AI knowledge apps have improved sharply from what he was able to build two months ago, but they still struggle on harder questions and remain slower and less effective than models that can run on a laptop. In a Sept. 26 post on X, Buterin said the weakest results came from travel-style queries. He tested the prompt asking for the best vegan restaurant in his current city, and said every app he tried performed poorly. The apps came out of a community-led bounty on decentralized bounty platform poidh, where contributors locked up a total of 1.098 ETH. Buterin himself supplied 1 ETH, more than 90% of the pool, even though the bounty description said it was independently created in response to his earlier post and that he would not judge entries unless he explicitly chose to do so. The challenge asks builders to create the best offline AI research app for Android under tight constraints: support for Android and GrapheneOS, a 12GB memory cap, a 50GB total limit for app, model and database, no network requests after installation, and open-source code on GitHub. As of Sept. 26, three submissions had been filed, and none had been publicly confirmed by Buterin as meeting the target.

2550
Vitalik says offline mobile AI apps have improved, but still trail laptop models
SemiAnalysis
Sep 22, 2026

SemiAnalysis says memory bandwidth matters more than capacity in AI inference, with scheduling emerging as a core layer

SemiAnalysis has published a report breaking down the underlying architecture of large-model inference services, arguing that the rise of mixture-of-experts, or MoE, models has turned inference into a multi-stage pipeline rather than a single compute task. The report describes that pipeline as consisting of Prefill, Midfill, Decode Attention, and Decode Experts, with each stage placing different demands on compute, memory bandwidth, and networking. Its central conclusion is that, in most inference scenarios, memory bandwidth carries more economic value than raw capacity. According to the report, high-bandwidth memory can improve token generation efficiency, while idle HBM mainly adds cost. SemiAnalysis also projects that by 2027, a single pipeline stage may require roughly 400 GB to 500 GB of local fast memory, though processed KV cache should be moved to CPU DRAM and lower-cost network storage to avoid tying up scarce HBM resources. The report also identifies the scheduling layer as a critical part of inference infrastructure. It says Prefill and Midfill are relatively predictable in runtime, while decode latency is more variable and can create backlog and delay-feedback oscillation. SemiAnalysis further compares integrated and disaggregated architectures, and includes simulator-based projections for Kimi K3 on Nvidia B200, B300, and GB200 systems.

1200
SemiAnalysis says memory bandwidth matters more than capacity in AI inference, with scheduling emerging as a core layer
Total 41 items
Go topage