GLM-5.3-Flash 공개, 저비용 컴퓨터 사용 모델과 외부 AI 연구 데이터 개방

GLM-5.3-Flash scores 57 on the Artificial Analysis Intelligence Index. At $0.09 Cost per Task, it sits comfortably on the Intelligence vs. Cost per Task Pareto frontier @Zai_org has released GLM-5.3-Flash, a smaller and cheaper sibling to GLM-5.3 at 320B total parameters and just 18B active parameters. GLM-5.3-Flash supports low/high/max reasoning efforts, and scores 57 evaluation on the Artificial Analysis Intelligence Index with max reasoning effort. This places the model only 3 points behind GLM-5.3 at 60 and in line with GPT-5.6 Terra and Muse Spark 1.2. On Z AI's first-party API, GLM-5.3-Flash is priced at $0.15 / 1M input tokens and $0.50 / 1M output tokens, just over 10% of the price of GLM-5.3. Cached input tokens are priced at $0.026 / 1M tokens, an 80% discount. Its Cost per Task on the Intelligence Index is $0.09, compared to $0.68 for GLM-5.3 (max), and it sits on the Pareto frontier for Intelligence vs. Cost per Task. Key results: ➤ GLM-5.3-Flash is 3 points behind GLM-5.3 (max) on the Artificial Analysis Intelligence Index, at ~7.5x lower Cost per Task. At $0.09 per Intelligence Index task against $0.68 for GLM-5.3, it sits on the Pareto frontier for Intelligence vs. Cost per Task. It ties GPT-5.6 Terra ($0.51) and Muse Spark 1.2 ($0.40) at 57 while costing ~5.7x and ~4.4x less per task. ➤ GLM-5.3-Flash is less token efficient, but its low per-token pricing means this does not translate into a high Cost per Task. The model used 149M output tokens to run the Intelligence Index, ~11% fewer than GLM-5.3 at 168M, but more than Kimi K3 (133M) and Qwen3.8 2.4T A95B (136M) which score the same on the Intelligence Index. Reasoning tokens account for 134M of the 149M total (~90%). ➤ GLM-5.3-Flash matches GLM-5.3 on real-world agentic work on GDPval-AA v2. With an Elo of 1770, the model is tied within the margin of error for GLM-5.3 and Grok 4.6. This places it behind only Claude Opus 5 (xhigh and max). On Terminal-Bench v2.1 it also matches GLM-5.3 (84.3% vs 83.9%), and on τ³-Banking it trails by 3.1 p.p. at 47.2%. ➤ GLM-5.3-Flash demonstrates good real-world knowledge and hallucination rate, scoring +7 on AA-Omniscience. Its AA-Omniscience Accuracy is 28%, 6 p.p. below GLM-5.3 (max) at 34% and well below GPT-5.6 Terra at 47%. However, with a Hallucination Rate of 28%, it is an improvement over GLM-5.3 at 30%. In real-world knowledge, GLM-5.3-Flash knows less than the bigger models and frontier proprietary models in its Intelligence Index tier with an accuracy of 28%. Additional model details: ➤ Pricing: On Z AI's first-party API, $0.15 / 1M input tokens and $0.50 / 1M output tokens . Cached input tokens are priced at $0.03/ 1M tokens, an 80% discount. ➤ Accessibility: Accessible through Z AI's first-party API at launch. ➤ Size: 320B total parameters with 18B active parameters ➤ License: MIT ➤ Context Window: 400k

Introducing GLM-5.3-Flash - Leading capabilities at a highly competitive price - Natively multimodal with a 1M-token context window - A 320B-A18B model released under the MIT License - Previously previewed as Ox Alpha, running entirely on Chinese AI chips Blog: http:// z.ai/blog/glm-5.3-f lash … Available now across all official platforms: Weights: http:// huggingface.co/zai-org/GLM-5. 3-Flash … API: http:// docs.z.ai/guides/llm/glm -5.3-flash … Coding Plan: http:// z.ai/subscribe ZCode: http:// zcode.z.ai/en Chat: http:// chat.z.ai AutoClaw: http:// autoclaw.z.ai

Qwen3.8-Flash can now be run locally! The 125B MoE model outperforms Claude-Opus-4.6 (Max). Run on 75GB RAM via Unsloth GGUFs. Qwen3.8-Flash-Next enables CPU RAM / unified mem setups to deliver near VRAM speeds. Guide: https:// unsloth.ai/docs/models/qw en3.8-next … GGUF: https:// huggingface.co/unsloth/Qwen3. 8-Flash-Next-GGUF …

API is live on QwenCloud: https:// qwencloud.com/models/qwen3.8 -flash … Let's build something!

Excited to share that Navigator n2 is officially out! The core idea behind n2 is simple: models should use computers the computer way, not the human way. There’s no single best interface. GUI, CLI, tools, code — pick the right one at every step. The hard part is making those choices well over long horizons: knowing when to switch, when not to, and how to keep making progress over hundreds of steps. The training loop ended up following the same idea: training computer-use models is itself a computer-use task. We use computer-use agents to explore applications, create tasks, test verifiers, find reward hacks, edge cases, and where the latest model fails. Those failures shape the next round of training data. It creates a pretty natural recursive loop: better models make better data, which makes better models. The result is a frontier-level model at a fraction of the cost— 65.2% on OSWorld 2.0 at $1.46 per task. Very proud of the team and where n2 landed!

Excited to share that Navigator n2 is officially out! The core idea behind n2 is simple: models should use computers the computer way, not the human way. There’s no single best interface. GUI, CLI, tools, code — pick the right one at every step. The hard part is making those choices well over long horizons: knowing when to switch, when not to, and how to keep making progress over hundreds of steps. The training loop ended up following the same idea: training computer-use models is itself a computer-use task. We use computer-use agents to explore applications, create tasks, test verifiers, find reward hacks, edge cases, and where the latest model fails. Those failures shape the next round of training data. It creates a pretty natural recursive loop: better models make better data, which makes better models. The result is a frontier-level model at a fraction of the cost— 65.2% on OSWorld 2.0 at $1.46 per task. Very proud of the team and where n2 landed!

Today we’re introducing Gemini 3.5 Transcribe, our latest transcription model built for incredibly precise, smart dictation across your favorite apps and devices. Remember when traditional speech-to-text meant shouting over background noise, constantly hitting backspace to fix misspelled words, and manually deleting every "um" and "uh"? Those days are over. Gemini 3.5 Transcribe isn't just dictation — it’s active intelligence with precise, context-aware speech-to-text support in 85+ languages. The model automatically filters out filler words, formats unstructured speech, and even pairs with your screen context to execute voice commands. Watch as Gemini 3.5 Transcribe removes filler words and uses multimodal capabilities to seamlessly turn messy voice input and local files into a polished email draft.

Say hello to Gemini 3.5 Transcribe! - Build apps that understand user speech / intent, even w/ multiple speakers! - Auto-detection of 85+ languages out of the box - Custom vocab adaptation for specialized jargon... SGTM:) API available now in @GoogleAIStudio and Gemini

For the first time, we’ve given external researchers a way to study AI’s impacts using real, privacy-preserved Claude usage data. To date, this work has only been possible within AI labs. We can’t tell the whole story alone, so we opened up our tools.

Ever since co-founding Anthropic I've been obsessed with the need for third-party measurement of AI systems. (It's in my tweet announcing Anthropic: https:// x.com/jackclarkSF/st atus/1398304973205630991 …). Now, we're piloting a way for outside researchers and organizations to study what's happening on AI platforms like Claude while protecting user privacy. This matters because AI systems are giant, complex sociotechnical things with innumerable properties. To think AI labs are going to be able to figure out all the appropriate ways to measure and assess these systems is hubristic and just obviously wrong. So we need to figure out ways to externalize not only the properties of the systems, but also data about how these systems are interacting with the people in the world, which is platform telemetry. What we're trying to do here is prototype the next form of transparency that may be necessary, which is platform transparency while preserving user privacy. Along with this, I want to find ways to empower researchers outside Anthropic to perform their own studies on the kinds of data that companies uniquely have access to, as I think this is a good way to create a more vibrant external research ecosystem. This is another step the larger project for The Anthropic Institute about measuring and externalizing more details about the properties of our AI systems and the platforms we deploy them on, and sits alongside existing efforts from our Economics team via the Anthropic Economic Index. We are now soliciting research ideas from others for our next wave of projects - please apply!

Brain is our self-improving memory system for Perplexity Computer. It compiles sessions, files, and sources into a structured knowledge wiki. New evals build on our initial results, improving correctness by 9.3 points, currentness by 8.0, and recall by 8.9 with 15% fewer tokens.

Aravind Srinivas

@AravSrinivas

1일 전

Perplexity Computer for Max users comes with a background Dream agent that continually ingests context from files and connected apps and builds multi-hop context graphs in a perpetual compounding loop. Results show a significant jump in correctness, recall and token efficiency.

Thanks to @tomaarsen for letting me contribute a tiny experiment (the dashed triangles) to this yesterday. I wanted to demonstrate my claim that late interaction's far stronger quality does NOT generally require a "storage overhead", with technology (PLAID) that was released in 2022 before GPT-3.5. Extremely tiny 307M-parameter mLateOn models, even zero-shot off the shelf ones, have no trouble handily beating every single-vector method, including 26x bigger Qwen3-Embedding-8B by double-digit nDCG %. Moreover, Tom's lightly-finetuned mLateOn-medical can trivially do so while representing hundreds of millions of tokens with ~1 GiB, an index size that is SMALLER than Qwen3's single-vector fp16 representations*, which is almost certainly the way you're currently storing Qwen embeddings. *Good luck compressing single-vector representations with that fidelity esp if OOD.