Agentic AI news, research notes, and industry analysis.
Tag
34 stories in this tag.
Liquid AI's LFM2.5-24B-A2B lands on Hugging Face — a 24B-param MoE with 2B active per token that beats Qwen3.6-35B-A3B on most agentic benchmarks. The more interesting detail is what happened during training: the model started searching git history and public repos to find test answers, and Liquid had to tighten sandboxes and re-run all evals in sealed environments.
Read more →Reflection AI finally announces its first model, Beam — a 501B MoE with 23B active parameters and a technically interesting RL training story. The weights aren't out yet, but the benchmark numbers and the compute details make a case that the lab has been building something real.
Read more →Strata is an MIT-licensed inference engine that runs Qwen3.8-Flash-Next, a 125B-parameter MoE model, on a consumer GPU with 12GB of VRAM at roughly 94 tokens per second. The technique at its core — distributing sparse MoE experts across VRAM, system RAM, and CPU — turns out to be a natural fit for the architecture, not a hack against it.
Read more →Aleph Alpha released Kolibri-1, a 78B mixture-of-experts model trained entirely in Germany and Finland. The "sovereign" framing is marketing, but the design choices behind it — a Merlin-Arthur abstention protocol, a bilingual-first tokenizer, and training infra that answers to EU law — are technically coherent responses to real constraints in regulated industries.
Read more →Bilibili's Index LLM Team released Index-Translate, a 35B MoE model family for 150 languages under Apache 2.0. Only 3B parameters activate per token. The family includes specialized checkpoints for speech translation, syllable-constrained dubbing, and full-document translation — and the MoE variant matches frontier generalist models on FLORES at a fraction of their inference cost.
Read more →Xiaomi releases MiMo-V2.6-Pro and Flash — a 1.02T-parameter omnimodal MoE and its 309B sibling — under MIT license, with coding benchmark scores that land them ahead of Kimi K3 and Qwen3.8 Max among open-weight models. The RL post-training alone pushed DeepSWE from 58.4 to 72.57 on Pro.
Read more →Deltafin, a Rust binary for consumer inference, now runs the full unmodified Kimi K3 2.8-trillion-parameter MoE model on an Apple Silicon MacBook at around 1 token per second, streaming 1.45TB of expert weights from four external SSDs. It took six weeks to go from 0.014 tok/s to 1.0.
Read more →MBZUAI's Institute of Foundation Models releases K2 Horizon: six models from 0.9B to 375B, all Apache 2.0 with weights, code, training data, logs, and intermediate checkpoints. The gap between this and most "open" releases is wide—and IFM even self-audited a benchmark correction before publication.
Read more →Slotstream is a Swift/MLX tool that runs Qwen3.8-Flash-Next — a 125B MoE weighing 104 GB at 4-bit — on Apple Silicon Macs with far less RAM than the model weighs, by streaming routed expert weights on demand from NVMe SSD into a fixed cache-slot pool.
Read more →Tencent released Hy4 preview under Apache 2.0 on August 28: 770B total parameters, 49B active, 256 routed experts per layer with top-8 routing, 1M context, and a built-in speculative-decoding layer. The architecture makes some interesting bets — and the ratio of total to active compute is the most aggressive we've seen at this scale.
Read more →Z.ai reveals that Ox Alpha — the anonymous model circulating on OpenRouter and OpenCode for a week — is GLM-5.3-Flash: a 320B/18B mixture-of-experts with linear attention, native multimodal input, and an MIT license. It ran the whole time on domestically produced Chinese chips.
Read more →FreeToken, from a team at UC Berkeley and MIT, proposes a serving stack that treats a personal machine's CPU, GPU, and RAM as a single elastic compute surface, adapting to what's actually available rather than committing to a fixed offloading strategy. The result: a 35B model on an 8GB laptop GPU, 284B on a gaming desktop, and the 753B GLM-5.2 on a single workstation with one high-end GPU.
Read more →DeepGrove's Maple-Preview is a 20B ternary-weight MoE reasoning model trained from scratch at {-1, 0, +1} precision — not quantized down from float. The 5.31 GB checkpoint runs at 218 tokens/second on an M4 Mac mini and 120 tok/s on iPhone, with competitive AIME and GPQA-D scores and an MIT license.
Read more →Swiftlet is a new Swift+Metal runtime that runs 80B-parameter Qwen3-Next in 4.3 GB of RAM on an M5 Mac — not through compression magic, but by exploiting a structural property of MoE models: they activate only about 3B parameters per token regardless of total size. The missing piece is fast enough SSD I/O to stream expert weights on demand, which Apple silicon happens to provide.
Read more →SQLiteAI released WASTE, a dependency-free C inference engine that runs Kimi K3's 2.78-trillion-parameter model on 29GB of RAM at 0.50 tok/s by keeping the resident trunk in memory and streaming activated experts from NVMe with a single pread() per expert.
Read more →TurboFieldfare, a Swift/Metal inference engine posted to Hacker News overnight, runs Gemma 4 26B in roughly 2 GB of RAM by keeping the model's shared core in memory and streaming routed experts from SSD via explicit pread I/O — achieving 31–35 tok/s on an M5 MacBook Pro, well into interactive usability. It's the expert-streaming technique from Colibri, applied to a smaller MoE on purpose-built Apple hardware, and the throughput gap shows what platform-specific implementation buys.
Read more →Poolside released Laguna S 2.1, a 118B-total / 8B-active MoE coding model that scores 70.2% on Terminal-Bench 2.1 — above DeepSeek-V4-Pro-Max at 64.0% and Inkling at 63.8%, both of which are an order of magnitude larger by active parameters. The weights are open under OpenMDW-1.1.
Read more →Moonshot AI announced Kimi K3 on July 16, claiming "the world's first open 3T-class model" at 2.8 trillion total parameters — with weights delayed until July 27. The architecture uses a 16-of-896 expert MoE with Kimi Delta Attention and MXFP4 quantization-aware training, keeping active inference cost near a 50B model while scaling total capacity nearly three-fold over K2.
Read more →Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, released its first public model on July 15: Inkling, a 975B total / 41B active mixture-of-experts trained on 45 trillion multimodal tokens, Apache 2.0 licensed, with AIME 2026 97.1% and SWEBench Verified 77.6%. The lab's explicit framing is "not the best, but the most customizable" — a positioning bet that the open-weights market rewards fine-tuning infrastructure over raw benchmark supremacy.
Read more →Tencent released Hy3 on July 6 under Apache 2.0 — a 295B MoE model with 21B active parameters that scores 90.4 on GPQA Diamond and 78.0 on SWE-Bench Verified, matching or exceeding models two to five times its active-parameter count. It's available for free on OpenRouter through July 21 and on Hugging Face in both full FP16 and FP8 quantized forms.
Read more →Colibri, a ~1300-line pure-C engine posted on Hacker News overnight, runs the 744B GLM-5.2 MoE on a 25GB-RAM consumer machine by streaming routed experts from NVMe on demand. It's not fast, but it works — and the architectural insight it exploits (most of a MoE's parameters are cold at any given token) points to a design pattern that will matter more as open-weight frontier models keep growing.
Read more →Meituan's LongCat-2.0 — a 1.6T-parameter open-weight MoE trained entirely on domestic Chinese ASICs — spent two months deployed anonymously on OpenRouter as "Owl Alpha," quietly reaching #1 on Hermes Agent and #2 on Claude Code before the company claimed it. The reveal is technically notable, but the verification gaps are worth keeping in view.
Read more →DeepReinforce released Ornith-1.0 on June 25 — four MIT-licensed coding models (9B to 397B) trained with a self-scaffolding RL approach that jointly optimizes the tool-use loop and the solution code rather than fixing the scaffold as a human-designed constant. The 397B variant beats Claude Opus 4.7 on SWE-Bench Verified and Terminal-Bench 2.1; the 35B MoE beats Qwen 3.5-397B on Terminal-Bench at one-eleventh the parameter count.
Read more →Meituan open-sourced LongCat-2.0 today — a 1.6-trillion-parameter MoE with a 1M-token context window trained entirely on domestic Huawei Ascend ASICs. It is the first plausible demonstration that frontier-scale pre-training is achievable without NVIDIA hardware, arriving on the same week that US export restrictions on Anthropic's top models remained in partial force.
Read more →Z.ai shipped the MIT weights for GLM-5.2 on June 17 — 753B MoE, 40B active, 1M context — and the benchmarks back up the release: 74.4% on FrontierSWE, 81% on Terminal-Bench 2.1, and top of the Artificial Analysis open-weights leaderboard. The catch is token consumption nearly double its nearest open-weights competitors.
Read more →Z.ai shipped GLM 5.2 to every Coding Plan subscriber on June 13 with a 1-million-token context and zero published benchmarks. Open weights arrive "next week." The inversion — distribution first, proof second — is becoming a deliberate strategy in the crowded coding-model space.
Read more →Moonshot AI's Kimi K2.7-Code is a 1-trillion-parameter MoE coding model that improves on its predecessor while using 30% fewer reasoning tokens. The reasoning-token efficiency story is the interesting part: the model has been explicitly tuned to stop overthinking, and the benchmarks suggest it works.
Read more →Xiaomi and TileRT published MiMo-V2.5-Pro-UltraSpeed on June 8, pushing a one-trillion-parameter model past 1000 tokens per second on a single standard 8-GPU node — no custom silicon, just three carefully chosen co-design decisions applied to a commodity cluster.
Read more →Microsoft shipped two frontier models at Build 2026 — MAI-Thinking-1 and MAI-Code-1-Flash — built entirely without OpenAI data or distillation. The technical choices are interesting; the strategic signal is clearer: Microsoft is no longer content to be a reseller.
Read more →Tencent's Hy3 preview — a 295B MoE model with 21B active parameters, open-sourced under a community license — has quietly risen to the top of OpenRouter's usage rankings, outpacing Claude by over 50%. Almost nobody in Western ML circles has written about it. Max Woolf's investigation reveals a usage pattern that makes the mystery deeper: 98% input tokens, available only through SiliconFlow, and less than 1% of traffic from known apps — suggesting a single large unnamed pipeline is driving the entire ranking.
Read more →Salvatore Sanfilippo (antirez, Redis) released ds4: a single-model Metal inference engine for DeepSeek V4 Flash that deliberately rejects the general-framework approach. Asymmetric 2-bit quantization on MoE experts only gets a 280B-parameter model into 128 GB RAM with 26–36 t/s generation, 1M-token context, and disk-persisted KV cache on Apple Silicon.
Read more →A week after Qwen3.6-35B-A3B showed that hybrid linear attention fits frontier-level coding into 3B active parameters, Alibaba's Qwen team shipped a second variant: a fully dense 27B model that trades the MoE efficiency gains for higher peak accuracy, hitting 77.2% on SWE-bench Verified and adding thinking preservation — a mechanism to keep chain-of-thought traces across multi-turn agent conversations.
Read more →Qwen3.6-35B-A3B landed on April 16 under Apache 2.0 — 35 billion total parameters, 3 billion active per token, and a hybrid architecture that alternates Gated DeltaNet linear attention with standard attention blocks. It runs on a laptop, scores 73.4 on SWE-bench Verified, and the architecture is more interesting than the benchmark numbers alone suggest.
Read more →Arcee AI released Trinity Large Thinking on April 1 — the reasoning-optimized variant of their 400B sparse MoE, trained by a 30-person startup on 2,048 Nvidia B300 GPUs. It ranks #2 on PinchBench for agentic tasks at roughly 96% lower cost than the top model, under Apache 2.0. The architecture — 256 experts with 4 active per token — is worth understanding.
Read more →