The story so far
Weights, every dated article, and repo commits on one rail.
2 dated articles and 12 repo commits so far
Teal is the weights (2025-06-28), amber triangles are 2 dated articles (each links to it), grey ticks are repo commits.
GLM-4.1V-9B-Thinking
10.3B parameters, images and text in, text out, 64K tokens of context.
Verified 2026-10-25 Rank #268 in LLMs · score 84.9 method v1 Sources 3 Downloads 30d 88.7k · #328 of 619 tracked
Do people use it?88.7kdownloads, 30 days#328 of 619 we track · below the median+0 likes, downloads −33% this week791 likesHugging Face2026-10-24
Where it's used
Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.
No API provider serves it, so its downloads are people running it themselves: 12 quantized builds and 33 Spaces.
- As a service0no API provider listed
- On own hardware12quantized builds
- Built on1513 fine-tunes · 2 adapters · 0 merges
- In apps33Hugging Face Spaces
Top Spaces running it: prithivMLmods/Multimodal-VLM-v1.0 ♥ 34 · mgbam/builder ♥ 5 · Kakashi-hatake/GLM-4.1V-9B-Thinking ♥ 1 · salemshilin/dalin-video-prompt-generator ♥ 1
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits a 16 GB card at INT4
About 6.2 GB at INT4, 12 GB at FP8, 25 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
What people built on it
27 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
27 derivatives on the Hub, led by 13 fine-tunes
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Reference
Everything else, folded: the full tables, every source.
What the card says
2 cited sentences
Summary generated, cited
GLM-4.1V-9B-Thinking is a 10.3B parameter image-text-to-text model based on the GLM-4-9B-0414 foundation model.[1][2] It utilizes the Glm4vForConditionalGeneration architecture and supports a 64k context length.[3][4]
- parameters human
- “Based on the [GLM-4-9B-0414](https://github.com/zai-org/G…”
- architectures
- “Supports **64k** context length.”
Generated 2026-10-07 15:04 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.
Use-case evidence
4 use cases
What it is good at 4 use cases
Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.
- Understand images and videostructuraldecision-model estimate p=1.00tag: image-text-to-textpipeline: image-text-to-text+1 more
- Reason through hard problemsstructuraldecision-model estimate p=1.00name: Thinkingtag: reasoning
- Chat and follow instructionsstructuraldecision-model estimate p=0.70tag: conversationalchat template present
- Run on a laptop or phonestructuraldecision-model estimate p=0.83
Facts
22 fields
Listed in LLMs
Facts
- GLM-4.1V-9B-Thinking has 10.3B parameters. (safetensors metadata, observed 2026-10-24)
- Context length is 65,536 tokens (64K). (config.json max_position_embeddings, observed 2026-10-24)
- License: MIT (friction: none). Permissive. (license file, observed 2026-10-24)
- Created 2025-06-28, last modified 2026-07-22. (HF API, observed 2026-10-24)
| Pipeline tag | image-text-to-text |
|---|---|
| Library | transformers |
| Architectures | Glm4vForConditionalGeneration |
| Model type | glm4v |
| Dtype | bfloat16 |
| Params by dtype | BF16: 10,292,777,472 |
| Size band | L |
| MoE | no |
| Context length | 65,536 (config.json max_position_embeddings) |
| Languages | en, zh |
| Papers | arXiv:2507.01006 |
| License | MIT (mit) |
| Hub snapshot | 2026-10-25 |
Servability
vLLM · SGLang
Servability
vLLM: supported · Glm4vForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Glm4vForConditionalGeneration · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed
vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.
Hardware fit
3 estimates
Hardware fit ESTIMATE
Weights-only memory for GLM-4.1V-9B-Thinking: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 25 GB · Single high-end GPU (<80 GB) |
|---|---|
| FP8 (1 byte/param) | 12 GB · Single GPU (<24 GB) |
| INT4 (0.5 bytes/param) | 6.2 GB · Single GPU (<24 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.
All variants
27 variants
Lineage
Base model: zai-org/GLM-4-9B-0414
quantized: 12finetune: 13adapter: 2
Top variants (27 shown, ranked by downloads)
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| unsloth/GLM-4.1V-9B-Thinking-GGUF | quantized | GGUF | — | — | mit | unsloth | 46 | 103.9k |
| tiny-random/glm-4v | finetune | — | 9.99M | S | — | tiny-random | 0 | 10.4k |
| mradermacher/GLM-4.1V-9B-Thinking-i1-GGUF | quantized | GGUF | — | — | mit | mradermacher | 1 | 8.5k |
| DevQuasar/zai-org.GLM-4.1V-9B-Thinking-GGUF | quantized | GGUF | — | — | — | DevQuasar | 1 | 7k |
| mradermacher/GLM-4.1V-9B-Thinking-GGUF | quantized | GGUF | — | — | mit | mradermacher | 2 | 6.9k |
| airtightsquid/TA_GLM4.5_lora_finetune | adapter | — | — | — | — | airtightsquid | 1 | 5.2k |
| Rainnighttram/GLM-4.1V-9B-Thinking-bnb-4bit | quantized | -bnb | 10.3B | L | — | Rainnighttram | 13 | 2.8k |
| unsloth/GLM-4.1V-9B-Thinking | finetune | — | 10.3B | L | mit | unsloth | 2 | 2.6k |
| mlx-community/GLM-4.1V-9B-Thinking-8bit | quantized | -8bit | 10.3B | L | mit | mlx-community | 3 | 940 |
| yujiepan/glm-4.1v-tiny-random | finetune | — | 9.99M | S | — | yujiepan | 0 | 944 |
| yujiepan/glm-4v-tiny-random | finetune | — | 9.99M | S | — | yujiepan | 0 | 929 |
| MrMeOrYou/GLM-4.1V-9B-Thinking-Q4_K_M-GGUF | quantized | GGUF | — | — | mit | MrMeOrYou | 0 | 473 |
| turboderp/GLM-4.1V-9B-Thinking-exl3 | quantized | exl3 | — | — | mit | turboderp | 5 | 103 |
| NalbertLeal/GLM-4.1V-9B-Thinking-Q4_K_M-GGUF | quantized | GGUF | — | — | mit | NalbertLeal | 0 | 273 |
| AINovice2005/quantized-GLM-4.1V-9B-Thinking | quantized | — | — | — | — | AINovice2005 | 0 | 234 |
| vvpml/GLM-4.1V-9B-Thinking-Q6_K-GGUF | quantized | Q6_K | — | — | mit | vvpml | 0 | 234 |
| tiny-random/glm-4.1v | finetune | — | 9.99M | S | — | tiny-random | 1 | 173 |
| Bickleigh/GLM-4.1V-9B-Thinking-Q4_K_M-GGUF | quantized | GGUF | — | — | mit | Bickleigh | 0 | 161 |
| amazingvince/geom3k_CoT_sft_5e5 | finetune | — | 10.3B | L | — | amazingvince | 0 | 102 |
| amazingvince/geom3k_CoT_sft | finetune | — | 10.3B | L | — | amazingvince | 0 | 90 |
| sasa2000/glm-4-1v-9b-thinking-text-only | finetune | — | 9.40B | L | mit | sasa2000 | 0 | 83 |
| amazingvince/geom3k_CoT_sft_5e5_lora | finetune | — | 10.3B | L | — | amazingvince | 0 | 58 |
| zeromodels/glm-4.1v-9b-thinking | finetune | — | — | — | mit | zeromodels | 0 | 33 |
| stakmodsco/Leadmindai | adapter | — | — | — | mit | stakmodsco | 0 | 0 |
| jfang/glv4-1v-grpo-all-linear-train700 | finetune | — | — | — | — | jfang | 0 | 0 |
| jfang/glv4-1v-all-linear-500sample-random | finetune | — | — | — | — | jfang | 0 | 0 |
| jfang/glv4-1v-all-linear-cn31-727 | finetune | — | — | — | — | jfang | 0 | 0 |
Repository changelog
12 commits
Repository changelog 12 commits
| Date | Commit | SHA |
|---|---|---|
| 2026-07-22 | Fix config model type (#22) | 3c1471e |
| 2025-10-25 | update for new format config of transformers lib / sglang lib | 9e9a4c5 |
| 2025-08-28 | Merge branch 'main' of hf.co:THUDM/GLM-4.1V-9B-Thinking | 17193d2 |
| 2025-08-28 | update | 99b56d9 |
| 2025-07-08 | Add notebook (#10) | b627c82 |
| 2025-07-03 | set 64000 to 65536 | def1e44 |
| 2025-07-02 | update | 3e29ece |
| 2025-07-02 | update | 3e17d8b |
| 2025-07-02 | update | 02437f7 |
| 2025-07-01 | readme | 145d799 |
| 2025-06-30 | init | a3ebdb1 |
| 2025-06-28 | initial commit | dd22be8 |
Articles and coverage
7 pieces
Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Official1
intl.cloud.baidu.com752 words
GLM-4.1V-9B-Thinking is a compact multimodal reasoning model that outperforms larger 72B-parameter competitors across 28 benchmarks through a three-stage pipeline combining a modified Vision Transformer, an MLP reasoning layer, and an autoregressive Transformer. The model achieves a 12.7% accuracy improvement in localization and a 9.3% F1-score gain in reasoning tasks while demonstrating an 87% reduction in GPU memory usage, a 72% decrease in response latency, and 91% lower energy consumption during inference compared to larger models.written, reviewed, cited
- The model's long-video understanding is limited to processing videos that do not exceed 5 minutes in duration.
- Adaptive Curriculum Learning evaluates model performance across 12 dimensions every 1,000 training steps to adjust difficulty.
- Few-shot adaptation for new domains requires substantial fine-tuning data rather than working out-of-the-box.
Analysis6
arxiv.orgauthorityabout the GLM 4.1V family13,810 words
GLM-4.1V-9B-Thinking achieves 85.8 on MMBenchV11-EN, 72.9 on MMStar, and 87.9 on AI2D. The model outperforms Qwen2.5-VL-7B on nearly all tasks across 28 public benchmarks.generated, cited
github.comauthorityabout the GLM 4.1V family1,251 words
Zhipu released GLM-4.1V-9B-Thinking, an open-source vision-language model designed to explore reasoning limits. It uses a thinking paradigm and reinforcement learning to match or surpass larger models on 18 benchmark tasks. The model supports 64k context length and is licensed under the MIT License.written, reviewed, cited
- The model supports a maximum of 1 video or 300 images per input but does not support simultaneous image and video inputs.
- Fine-tuning with Zero2 may result in zero loss, so Zero3 is recommended for stable training.
deepwiki.com
researchgate.netabout the GLM 4.1V family
liner.com2025-07-01about the GLM 4.1V family998 words
ai-tldr.dev2025-07-01727 words
Papers
1 paper
Papers 1
arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).
Similar models
3 models
Similar models
- Qwen3.5-9B · 9.65B same size band (L), same task (image-text-to-text), more providers, permissive license
- GLM-4.6V-Flash · 10.3B same size band (L), same task (image-text-to-text), more providers, permissive license
- Qwen3-VL-8B-Instruct · 8.77B same task (image-text-to-text), more providers, permissive license
Sources
3 sources
Sources
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:36 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:04 UTC · Last verified 2026-10-25
Follow GLM-4.1V-9B-Thinking
Watch this model
- /feeds/models/zai-org--GLM-4.1V-9B-Thinking.xml(RSS 2.0) · every change to GLM-4.1V-9B-Thinking
- /feeds/families/zai-org--GLM-4.1V.xml(RSS 2.0) · every change across the GLM-4.1V family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.