Thread

Respond to this Pin

1GLM-4.5 and GLM-4.5-Air Released

Language model rubric2026 criteria

Estimate from the version's published specs, benchmarks and reviews, 0-5; scores are only comparable within one criteria year. Scored as launched (GLM-4.5 flagship, 28 Jul 2025; GLM-4.5-Air is the 106B/12B-active tier, $0.20/$1.10). Reasoning null: no Artificial Analysis Intelligence Index entry on the 2026 scale at launch and no restatement found; Z.ai's headline was 3rd of 13 models on a 12-benchmark average (63.2; Air 6th, 59.8), not comparable to the anchors. Coding null: no Terminal-Bench 4.0 (Z.ai reported SWE-bench Verified 64.2%, Terminal-Bench 2.0 not scored here). Context 128K tokens (Z.ai docs) -> 1. Price $0.60 in / $2.20 out per 1M (The Batch, Aug 2025; Z.ai pricing page) = $1.00 blended at 3:1 -> 5. Speed null: none published at launch. Modalities: text only -> 1. Reviews: Simon Willison (notes no independent benchmarks yet), The Batch, The Decoder, a UN University C3 hands-on; independent measurement at launch was thin.^[1]

Reviews

GLM-4.5: Reasoning, Coding, and Agentic Abililties · simonwillison.net · Jul 28, 2025Ivan Fioravanti shared a video of the mlx-community/GLM-4.5-Air-4bit quantized model running on a M4 Mac with 128GB of RAM, and it looks like a very strong contender for a local model that can write useful code.
GLM-4.5, an Open, Agentic Contender · deeplearning.ai · Aug 6, 2025On BrowseComp (web browsing with multi-step searches), GLM-4.5 (26.4 percent accuracy) outperformed Claude 4 Opus (18.8 percent accuracy) but trailed o3 (49.7 percent accuracy).
Zhipu AI's GLM-4.5 is yet another open-source Chinese LLM closing the gap with Western models · the-decoder.com · Aug 17, 2025GLM-4.5 placed third out of 13 large language models across 12 agent, reasoning, and coding benchmarks with 63.2 points; the resource-efficient GLM-4.5-Air took sixth place with 59.8 points.
GLM-4.5: The Open-Source Model That Challenges Proprietary AI Dominance · c3.unu.edu · Aug 4, 2025After testing it myself with a nearly 2000-line codebase, I can confirm that this model delivers on its promise of handling large contexts with remarkable performance.

Sentiment over timePositive

Sentiment

Sentiment over time
GLM-4.5: Reasoning, Coding, and Agentic Abililties · Jul 28, 2025	0.60
GLM-4.5: The Open-Source Model That Challenges Proprietary AI Dominance · Aug 4, 2025	0.70
GLM-4.5, an Open, Agentic Contender · Aug 6, 2025	0.30
Zhipu AI's GLM-4.5 is yet another open-source Chinese LLM closing the gap with Western models · Aug 17, 2025	0.50

Alternate names

GLM
Alias

User Photos

Photos from readers, kept apart from the pin's sourced media. Rate what you ate or saw, and on a restaurant pin tag the dishes right on your photo.

Comments

References 683% CONFIDENCEOverall confidence: 83%How well the pin's source and references back up its dates.Weighted average of how firmly 6 references, the source included, support the pin's start and end times; a reference counts half as much for every 180 days older than the newestShow all pins at 75% confidence or better

The first entry is always the pin's source. Overall confidence is a weighted average of how firmly each reference supports the start and end times used above; a reference counts half as much for every 180 days older than the newest.

[1]90%
docs.z.ai/guides/llm/glm-4.5
docs.z.ai· Posted Oct 12, 2026· Starts Jul 28, 2025· 36% of score
Z.ai's GLM-4.5 launch post is dated 28 July 2025 (as cited by Wikipedia^[2]), and the weights for both GLM-4.5 and GLM-4.5-Air were uploaded to Hugging Face^[6] that day (04:44 UTC and 09:14 UTC). The repositories were created earlier, on 20 July, and Epoch AI lists 3 August and 20 July respectively, which are not the public release.
[2]72%
Z.ai - Wikipedia
en.wikipedia.org· Added Oct 12, 2026· 36% of score
Cites Z.ai's "GLM-4.5: Reasoning, Coding, and Agentic Abilities" post dated 28 July 2025 and records the international rebrand to Z.ai around this release.^[1]
[3]82%
Zhipu AI's GLM-4.5 is yet another open-source Chinese LLM closing the gap with Western models - The Decoder
the-decoder.com· Published Aug 17, 2025· 7% of score
Independent report: deeper-not-wider architecture, third of 13 models on 12 benchmarks, TAU-Bench 70.1, AIME 24 91.0, SWE-bench Verified 64.2, and the companion GLM-4.5V.^[1]
[4]90%
GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models - arXiv
arxiv.org· Published Aug 8, 2025· 7% of score
The technical report (submitted 8 August 2025): a 355B total and 32B activated parameter MoE model with a hybrid reasoning method supporting thinking and direct-response modes.^[1]
[5]85%
Zhipu AI (Z.ai) Releases Open-Weights GLM-4.5 Models That Perform Comparably To the Latest from Claude and DeepSeek - The Batch
deeplearning.ai· Published Aug 6, 2025· 7% of score
Independent analysis: 128K context and 96K output, API prices of $0.60/$0.11/$2.20 and $0.20/$0.03/$1.10 per million tokens, 22T training tokens, BrowseComp 26.4 against Claude 4 Opus 18.8, SWE-bench Verified 64.2.^[1]

Suggest a correction

Something missing or wrong? Say it in your own words: a link that backs this pin up, a different start or end date and why, or a fact it lacks or gets wrong. The AI checks it against this pin's sources, searches for better ones, and adds any page that backs you up. The pin's own sources still count most. A picture that shows something else, or shows it badly, is looked at too, and moved down or replaced.

Zhipu AI, which started calling itself Z.ai abroad around this release, put GLM-4.5 and the smaller GLM-4.5-Air on Hugging Face and its API platform on 28 July 2025 under the MIT licence, following Moonshot's Kimi K2 and Alibaba's Qwen3-235B update^[2][5][6]
GLM-4.5 is a 355B-parameter mixture-of-experts model with 32B active; GLM-4.5-Air has 106B total and 12B active; both switch between a thinking mode for reasoning and tool use and an instant non-thinking mode^[1][6]
Across 12 agentic, reasoning and coding benchmarks Z.ai ranked GLM-4.5 third at 63.2 (behind o3 at 65.0 and Grok 4 at 63.6) and GLM-4.5-Air sixth at 59.8, ahead of Claude 4 Sonnet at 59.2 and Gemini 2.5 Pro at 58.8^[3][6]
Individual results included 26.4 on BrowseComp against Claude 4 Opus's 18.8, 70.1 on TAU-Bench, 91.0 on AIME 24 and 64.2 on SWE-bench Verified, below Claude 4 Sonnet's 70.4 and Kimi K2's 65.4^[3][5]
Z.ai stressed parameter efficiency: GLM-4.5 uses half of DeepSeek-R1's parameters and a third of Kimi K2's, favouring deeper rather than wider layers^[3]
The multimodal GLM-4.5V built on the Air model followed on 11 August 2025^[2][3]

Notable features

Text-only, 128K-token context and 96K-token maximum output; function calling and switchable reasoning^[1][5]
API prices at launch: GLM-4.5 $0.60 per million input tokens, $0.11 cached and $2.20 output; GLM-4.5-Air $0.20, $0.03 and $1.10^[5]
Open-sourced base models, hybrid reasoning models and FP8 versions of both sizes, with a technical report on arXiv^[4][6]
Pre-trained on roughly 22 to 23 trillion tokens, including about 15 trillion of general text followed by 7 trillion on code and reasoning, then distilled from three specialist copies into one model^[3][5]
Z.ai's own coding evaluation across 52 tasks gave GLM-4.5 an 80.8% win rate against Qwen3-Coder and 53.9% against Kimi K2^[5]

Benchmarks

Z.ai's 12-benchmark average from the maker's chart on docs.z.ai^[1]; the per-category charts are in the media.

Model	Average of 12 benchmarks
o3	65.0
Grok 4	63.6
GLM-4.5	63.2
Claude 4 Opus	60.9
o4-mini (high)	60.4
GLM-4.5-Air	59.8
Claude 4 Sonnet	59.2
Gemini 2.5 Pro	58.8
Qwen3-235B-Thinking-2507	57.2
DeepSeek-R1-0528	55.9
Kimi K2	53.1
GPT-4.1	48.7
DeepSeek-V3-0324	46.3

Suggest a correction

Comments

Tags

GLM-4.5 and GLM-4.5-Air Released

Notable features

Benchmarks

Thread