Z.ai Launches GLM-5.3-Flash With 1M-Token Context and Hybrid Attention Architecture
Z.ai has released GLM-5.3-Flash, a native multimodal model built for coding and long-horizon agent tasks, featuring a 1M-token context window and a hybrid sparse-linear attention architecture. The model is available via OpenRouter at a discounted $0.075/$0.25 per 1M tokens through September 2026.
GLM-5.3-Flash — Quick Specs
Z.ai has released GLM-5.3-Flash, a native multimodal model designed for efficient coding and long-horizon agent workloads. The model is now listed on OpenRouter with a 1 million token context window.
Architecture and Positioning
According to Z.ai, GLM-5.3-Flash uses a hybrid sparse and linear attention architecture intended to preserve accuracy over long-context inputs while cutting compute overhead relative to standard dense attention. The company positions the model for coding tasks and multi-step agent workflows that require sustained context tracking over extended sessions.
The model is described as "native multimodal," indicating it processes multiple input modalities directly rather than through bolted-on adapters, though Z.ai has not published a detailed breakdown of supported input types (text, image, audio, etc.) in the material reviewed.
Pricing and Availability
GLM-5.3-Flash is priced at:
- Input: $0.075 per 1M tokens (list price $0.15, currently 50% off)
- Output: $0.25 per 1M tokens (list price $0.50, currently 50% off)
- Cache read: $0.015–$0.03 per 1M tokens depending on provider
The 50% discount is available through the Z.ai provider on OpenRouter until September 9, 2026, at 16:00 UTC. Standard rates apply after that window closes.
The model officially released on August 26, 2026, and is currently served through OpenRouter with a reported P50 latency of 3.28 seconds and throughput of 27 tokens per second on the best-performing provider. OpenRouter lists 24-hour availability at 98.71% across the past three days of monitoring.
What We Don't Know
Z.ai has not disclosed a parameter count, training data cutoff date, or independent benchmark scores (MMLU, HumanEval, or similar) for GLM-5.3-Flash in the material reviewed. Claims about the hybrid attention architecture's compute savings and long-context accuracy come directly from Z.ai and have not been independently verified.
What This Means
GLM-5.3-Flash extends Z.ai's GLM line into a lower-cost, higher-throughput tier aimed squarely at coding assistants and agent frameworks that need to hold large amounts of context — codebases, tool logs, multi-turn plans — without paying dense-attention compute costs. At $0.075/$0.25 per 1M tokens, it undercuts many flagship multimodal models on price while matching or exceeding their context windows at 1M tokens.
The real test will be third-party benchmarking once independent evaluations surface, since Z.ai's efficiency and accuracy claims for the hybrid sparse-linear attention design are unverified. For now, the aggressive discount pricing and OpenRouter's multi-provider routing suggest Z.ai is pushing for rapid adoption among developers building coding agents, betting on volume before the promotional pricing window closes in September 2026.
Related Articles
MiniMax releases M3.1-Flash: 428B-parameter multimodal MoE with 1M-token context and weights on Hugging Face
MiniMax has published MiniMax-M3.1-Flash on Hugging Face: a native multimodal mixture-of-experts model with about 428B total parameters, 23B activated per token, and a context window of up to 1M tokens. It builds on MiniMax-M3 with sparse attention, quantization-aware training and DSpark speculative decoding.
Xiaomi open-sources Xiaomi-Lens-0, a 4B multimodal model that outputs pixel-level mask tokens mid-answer
Xiaomi's SeerRay-Lab has open-sourced Xiaomi-Lens-0, a 4B multimodal model built on Qwen3.5-4B. The model emits mask tokens for pixel-level image regions in the middle of an answer. The lab also released a benchmark called LensBench.
Block announces Dessa, a speech-to-speech model trained from scratch on human conversation
Block announced Dessa, a speech-to-speech AI model it says was trained "from scratch on nothing but human conversation." The team says it can run locally on a MacBook Pro M4 Max "or better," and early access will go to a small group of AI safety researchers.
Kimi K3.1 Pro scores 55 on Artificial Analysis Index, fourth place behind Claude Opus 5.5, with API prices cut by a thir
Moonshot AI has released Kimi K3.1 Pro, which scores 55 on the Artificial Analysis Intelligence Index, an 11-point gain over Kimi K3 and fourth place overall. Parameter count is unchanged at 2.8 trillion, while API prices fall by 33 percent. Moonshot plans to release the weights by November 4, 2026.
Comments
Loading...