Mixture of Experts

Summary: DeepSeek V4.1 Flash in 30 Seconds DeepSeek-V4.1-Flash shipped on September 10, 2026. License is MIT, weights are on Hugging Face. 552B backbone parameters, but only 8B active while reading input and 16B while generating output. That asymmetry is the whole point of the model. New architecture: Causal Encoder-Decoder (CED). 40 layers, 20 causal encoder plus 20 decoder. The decoder’s global KV cache is projected from the encoder’s final hidden states instead of from each decoder layer’s own. Result: a global KV cache of 890 bytes per token. One quarter of V4-Flash, and 1/437 of DeepSeek-V1. Natively multimodal: images go through DeepSeek-ViT, a vision encoder trained from scratch, from the very start of language-model pre-training. Pre-training corpus is 45T tokens. Beats Claude Opus 5 on several agentic benchmarks: 90.6 vs 89.1 on Terminal-Bench 2.1, 74.2 vs 74.0 on DeepSWE v1.1, 54.8 vs 50.3 on AutomationBench. Reasoning effort is a continuous 1-100 dial, not an on/off thinking toggle. API pricing is $0.15 in / $0.60 out per million tokens off-peak, double that at peak. Roughly one twentieth of Claude Opus 5’s output price. DeepSeek is retiring V4 Pro: from September 14, 2026, deepseek-v4-pro requests get routed to V4.1 Flash and billed at Flash rates. Everyone in the open-weight race this year is chasing the same two numbers: active parameters and KV cache. The first decides what each token costs, the second decides how much memory a long context eats. GLM-5.3-Flash answered with hybrid attention, Qwen3.8-Flash-Next with 6B active parameters. ...

Summary: GLM-5.3-Flash in 30 Seconds GLM-5.3-Flash is Z.ai’s (formerly Zhipu AI) new model, released 26 August 2026. It is the first natively multimodal member of the GLM-5 series: text and images go through the same model. It is 320 billion parameters, but only 18 billion run per token. Layer count is roughly half of GLM-4.5’s: 45 against 92. The licence is MIT. Most strong Chinese models this summer shipped under bespoke community licences; here there is no fine print to read before shipping a product. It beats GLM-5.2 by a wide margin on coding and agentic tests (63.4 against 46.2 on DeepSWE) and approaches Claude Opus 4.8 overall, at roughly one tenth of GLM-5.2’s price. API pricing is $0.15 in / $0.50 out per million tokens, or $0.075 and $0.25 with the 50% discount running until 9 September 2026. It is the first model in the series to use hybrid attention: linear attention carries local dependencies, sparse attention retrieves distant context. Against GLM-5.3 that is 3x less attention compute and a 4.4x smaller KV cache. Before launch it was tested anonymously as ox-alpha on OpenCode and OpenRouter, where it became the most used model of the week. All of that traffic was served on Chinese AI chips. There is one race in open-weight models this summer: producing the same intelligence with less compute. Alibaba’s Qwen3.8-Flash-Next beat its own 397B sibling with 6 billion active parameters. Z.ai’s answer is GLM-5.3-Flash: 320 billion total parameters with only 18 billion running per token, leaving GLM-5.2 behind at a tenth of the cost. ...

Summary: Qwen3.8-Flash-Next in 30 Seconds Qwen3.8-Flash-Next is Alibaba’s new AI model, out in August 2026. It reads text and images and writes text back, tuned for writing code and running multi-step work on your behalf. Anyone can download the model files, live since 24 August. The license is not fully permissive though, so read it before shipping it in a product. The real story: Alibaba shipped this as a dry run for the next big release, Qwen4. The identifier inside the model files literally says qwen4_exp, as in “Qwen4 experimental”. It is 125B parameters in size, but only 6B of them run for any given word. Big-model knowledge, small-model bill. It reads about 750,000 words in one go (1M tokens), and at that length it is 8x faster than its much larger sibling. Training it cost roughly 1/9 of the 397B Qwen3.7-Plus, and it still beats that model on coding and office work. It is cheap to run: $0.16 in / $0.47 out per million tokens. The flagship in the same family costs 12x that. When Alibaba shipped Qwen3-Next, the pitch was: this is not a finished product, it is next generation’s architecture released early so the community can poke at it. That architecture then carried the whole Qwen3.5 through Qwen3.8 line. ...

Summary: DeepSeek V4 Pro in 30 Seconds DeepSeek V4 Pro 0813 went generally available on August 13, 2026. No press release, no blog post, just one line in a changelog. 1.7 trillion parameters, mixture-of-experts architecture, 1M token context, 384K token output ceiling. The weights are on Hugging Face under the MIT license. You can download them, modify them and ship them in a commercial product. Official scores are bold: 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE. Independent testing is more restrained: 53 on the Artificial Analysis Intelligence Index. Pricing is $0.435 in and $0.87 out per million tokens. From August 16 the off-peak rate is half of that. The catch: its smaller sibling V4 Flash scores 52 on the same index at one third of the price. You are paying triple for one point. Model launches have settled into a ritual: a teaser video, a wall of benchmark charts, an excited founder post. DeepSeek skipped all of it. DeepSeek V4 Pro 0813 went generally available on August 13, 2026 with no announcement at all. One line landed in the API changelog and the weights appeared on Hugging Face. ...

Summary: Qwen3.8-Max in 30 Seconds Qwen3.8-Max is Alibaba’s new flagship model, made generally available on August 2, 2026. 2.4 trillion parameters, 95 billion active (MoE architecture). It reads text, images and video, and returns text. Context window is in the 1 million token class. On one task it ran 125 hours (about 5 days) with no human input, rebuilding an experiment from a machine learning paper from scratch, confirming its six findings, then inventing a method that beats the paper. It beats Claude Opus 4.8 on most coding and agent tests, trades blows with Claude Fable 5 and GPT-5.6 Sol, and falls behind on some. API pricing is $2 input / $6 output per million tokens. Repeated input costs $0.25. This is the first time Alibaba has open-weighted a Max-class model. The weights landed on Hugging Face on August 12, 2026, though under Alibaba’s own Qwen3.8-Max license rather than Apache 2.0. Two days later, on August 14, Qwen3.8-27B followed: a dense 27B model under Apache 2.0 that fits on a single GPU. Ask an AI to “rebuild the experiment in this paper, then improve on it” and it normally stalls after a few turns, waiting for you to step in and steer. ...