Topic
A neural network design that activates only a few specialized "expert" subnetworks per input, scaling parameter count without raising compute cost.
Reflection AI says Beam, its 501-billion-parameter open-weight model, needs three to four times less compute than comparable open models. Buyers have only the company's figures until the weights ship later in October.
Perspective Coverage
3 publishersReality
Reflection AI's 501-billion-parameter Beam beats GLM 5.2 on SWE-bench Pro and trails it on Terminal-Bench, by the company's own scores. Its claimed compute saving covers only part of a serving bill, so planning around it as a US-built open model has to wait for outside tests.
Reality
Aleph Alpha released Kolibri, an Apache 2.0 German-English model that activates 3.46B of its 78.1B parameters per token. Each token costs about as much compute as a small model, yet a team hosting it in Europe still has to fit every expert in memory.
Perspective Coverage
6 publishersReality
Strata runs a 125B-class Qwen model at about 100 tok/s on one RTX 4090 because its sparse MoE activates only about 6B parameters per token. The design moves the hardware bill to system RAM, and the 100 tok/s figure holds mainly for code and structured output.
Reality
Moonshot AI released open weights for Kimi K2.7-Code, a trillion-parameter coding model that activates 32 billion parameters per token. Its headline gains come from Moonshot's own benchmarks, so teams paying for proprietary agents have to measure it on their own code.
Reality
Salvatore Sanfilippo's open-source ds4 engine runs a short list of large open-weight models locally by compressing their routed experts to about two bits. Even compressed, the supported builds need high-memory Macs or GPU systems that most people do not already own.
Reality
vLLM 0.30.0 ignores the per-module rank and alpha patterns in PEFT LoRA adapters, and one test put the error against PEFT at 70 times the unpatterned baseline. The adapter loads without complaint, so only a config check or a side-by-side against PEFT will catch it.
Reality
Ai2 released Olmo-core 3, an open mixture-of-experts training stack it has benchmarked at over one trillion total parameters. Its speedups are measured against Ai2's own earlier FSDP code, so teams on Megatron-Core must run their own comparison.
Perspective Coverage
3 publishersReality
VIDRAFT says POCKET-Darwin-180B, a 111 GB 4-bit GGUF build of its 180B mixture-of-experts model, runs in llama.cpp on about $1,400 of consumer hardware. The accuracy evidence so far is one MMLU-Pro comparison that VIDRAFT reports itself.
Reality
Xiaomi's public dashboard put the MiMo 2.6 Pro reinforcement-learning run at $1.05 million after about 51 hours, roughly $20,500 an hour. Its restart notes and token count give other teams an all-in reference for pricing their own RL runs.
Reality
Fudan University researchers studied 700-plus task logs from the Atria Dawn build and found humans made 85.5 percent of method and parameter calls. For teams copying the setup, output is capped by how fast people can decide and review what agents produce.
Reality
Developer xbill9 rebuilt Google's QAT Gemma 4 26B as int4 and fit it on one TPU v6e with 53,888 tokens of KV cache, against 3,456 for RedHat's FP8 build. Throughput nearly doubles, with accuracy checked on one classification suite.
Reality
Moonshot published 2.8 trillion open weights. At four bits per parameter that is about 1.4TB resident before any cache, which rules out the eight-way H100 node most teams assume.
Reality
Hy4 preview carries 2.6 times Hy3's parameters and 3.9 times its context window. Tencent is giving the weights away. That means the build column in next quarter's model budget gets priced in accelerator memory rather than in tokens.
Reality
AWS says running DeepEP over its EFA network on EKS gives mixture-of-experts reinforcement learning 40% more throughput. Whether that reaches another cluster depends on how much of each training step goes to expert traffic between nodes.
Reality
V4.1-Flash retires the V4 Pro line and carries two active-parameter counts, 763B total with 8B on input tokens and 16B on output, so one sizing number no longer covers both phases of a request. Baseten had it running on day zero.
Perspective Coverage
3 publishersReality
Moonshot AI's Kimi Linear reports 75% less KV cache and 6.3x faster decoding than MLA at 1M tokens, using three linear layers per attention layer. The speedup falls to parity at 4k tokens, so the saving goes to traffic that runs at hundreds of thousands of tokens.
Reality
Two agentic post-training jobs, a 1.02-trillion-parameter model and a 309-billion cousin, ran with step time, reward and infrastructure error rates on the open web. Both stopped near step 30, and both models are still unreleased.
Reality
Flash's off-peak input price is under a quarter of what V4-Pro cost, and on DeepSeek's own table it beats the old Pro checkpoint on Terminal-Bench, but it scores 36.8 on Humanity's Last Exam and no V4.1-Pro has a date.
Reality
Artificial Analysis scores the new Xiaomi model first among open weights and twelve points behind Claude Opus 5.5, and the cheaper Flash tier is the one an operator should put in front of a real queue.
Reality
Build · September 22, 2026 · 1 publisher
Build · September 21, 2026 · 1 publisher
Build · September 20, 2026 · 2 publishers
Invest · September 20, 2026 · 1 publisher
Science · September 18, 2026 · 1 publisher
Build · September 17, 2026 · 1 publisher
Build · September 17, 2026 · 1 publisher
Product · September 17, 2026 · 1 publisher
Build · September 17, 2026 · 1 publisher
Build · September 16, 2026 · 1 publisher
Build · September 15, 2026 · 1 publisher
Build · September 14, 2026 · 1 publisher
Build · September 14, 2026 · 1 publisher
Build · September 14, 2026 · 1 publisher
Build · September 13, 2026 · 1 publisher
Leadership · September 12, 2026 · 1 publisher
Build · September 11, 2026 · 2 publishers
Build · September 11, 2026 · 1 publisher
Product · September 11, 2026 · 1 publisher
Build · September 10, 2026 · 1 publisher
Build · September 9, 2026 · 1 publisher
Leadership · September 8, 2026 · 1 publisher
Build · September 8, 2026 · 1 publisher
Build · September 7, 2026 · 1 publisher
Build · August 28, 2026 · 5 publishers
Leadership · September 5, 2026 · 1 publisher
Build · September 4, 2026 · 1 publisher
Build · September 1, 2026 · 1 publisher
Build · August 31, 2026 · 1 publisher
Build · August 21, 2026 · 1 publisher