GLM 5.3 Flash

FelloAI's supplied excerpt confirms Z.ai released GLM 5.3 Flash on August 26, 2026, as a 320-billion-parameter mixture-of-experts that activates 18 billion parameters per token, the first natively multimodal model in the GLM-5 series with a 1M-token context and a hybrid sparse-and-linear attention design. MIT-licensed The supplied excerpt emphasizes that "Flash" is not a distilled or trimmed version of the GLM 5.3 flagship but a different model built on a newly trained base rather than post-trained on GLM 5.2's weights, and it reports that the GLM 5.3 flagship's promised 744B open weights still had not shipped at launch. The article

Linas's Substack recounts how GLM-5.3-Flash ran anonymously on OpenRouter and OpenCode as "Ox Alpha" for six days starting August 20, 2026, processing approximately 23 trillion tokens to become the most-used model on the platform at roughly 2.3x the volume of the next model. On August 26, Bloomberg confirmed it as Zhip The narrative highlights that Zhipu's Hong Kong shares closed more than 12% higher the following day (roughly 10x January's IPO price) and notes Stripe's agreement to acquire OpenRouter coincided with the mystery model's appearance. The piece emphasizes the deployment significance of running without NVIDIA chips and po

Yotta Labs' comparison piece documents that Z.ai shipped GLM 5.3 on August 14, 2026, as an API-only flagship built on a 320B-total / 18B-active MoE scaled from the GLM 5.2 base, followed by GLM 5.3 Flash on August 26 with MIT weights, native image and video input, and launch pricing around a tenth of the flagship's. Th The only shared benchmark between both launch tables is DeepSWE v1.1, where GLM 5.3 Flash scores 63.4 versus the flagship's 66.9, a gap of 3.5 points at one-tenth the price with open weights. Flash launched with a week of stealth usage as "Ox Alpha" before Z.ai claimed it, while the flagship launched benchmark-table-fi

The Medium/CodeToDeploy explainer summarizes GLM-5.3-Flash as released by Z.ai on August 26, 2026, with 320 billion total parameters and approximately 18 billion active parameters, a 1-million-token context window, and multimodal inputs including image and video. It positions the model as the newest member of the GLM-5 The article frames the release within a broader industry shift toward maximizing intelligence per unit of compute rather than sheer parameter count. It emphasizes the model's agentic coding capabilities and dramatically lower inference costs relative to frontier models. As a member-only Medium piece, verification is li

Ampere.sh's comparison reports Artificial Analysis Intelligence Index scores of 60 for GLM 5.3 and 57 for GLM 5.3 Flash, with Flash costing approximately one-ninth as much at normal API pricing ($0.15/M input, $0.50/M output vs. $1.40/$4.40). Flash activates only 18B parameters versus 40B for the flagship, but counteri The piece recommends GLM 5.3 Flash for most high-volume applications and multimodal agents, and GLM 5.3 for maximum coding quality and generation speed. Flash offers 3x usable quota on the Z.ai Coding Plan versus the flagship's 1x reference, and ships with public MIT weights while the flagship's 753B weights are "comin

DataCamp's supplied excerpt confirms GLM-5.3-Flash as Z.ai's cost-optimized model in the GLM-5.3 family, using a 320B-A18B MoE design with a 1M-token context window and native multimodal input. It scores 84.3 on Terminal-Bench 2.1, placing it within striking distance of Claude Opus 4.8 (85.0) and behind GPT-5.6 Terra ( The supplied excerpt reports a blended price of roughly $0.10 per million tokens (about one-tenth of GLM-5.3's $0.90 blended rate), output throughput of approximately 49 tokens per second, and a 306 GiB FP8 checkpoint that the article flags as ruling out lightweight local hosting. The piece also notes that Qwen3.8-Flas

LLM-Stats ranks GLM-5.3-Flash at composite rank 18 with capability tiers showing Top 10% in Chat (3/117) and Tool Calling (7/194), Top half in Vision (24/208), Coding (32/267), and Reasoning (17/363), with lower placements in Math (38/327), Legal (70/211), and Finance (86/228). Blended price is reported at $0.17/M agai A quality-tracker sigma timeline shows a +8.43σ baseline shift and subsequent improvements (+0.89σ at 7d, +0.79σ at 30d), suggesting post-launch quality gains. Sources for benchmark scores are attributed to Artificial Analysis and z.ai, providing a link to first-party data even though the leaderboard itself is third-pa

BenchLM's aggregator page for GLM-5.3-Flash reports a field-median capability score of 66.1/100 against a 56.3 median, with an advertised 1M-token context window. Verified benchmark coverage includes 6/6 Agentic (score 60.2, rank 19/154, 88th percentile), 6/6 Coding (58.1, rank 28/154, 82nd percentile), 5/5 Multimodal Self-hosting is the listed pricing mode with an infrastructure-cost input median around $0.95 and speed not measured. The page frames the model as self-hostable with FP8 inference implied by its inclusion in a benchmark catalog, and it carries no pricing or availability claims for any specific serving provider, keeping