# GLM-4.5 On-Prem Sizing Calculator: Flagship and Air Variant Hardware Plan

> Source: https://netray.co/tools/glm-4-on-prem-sizing-calculator
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: Free Tools

## Short answer

Size VRAM, GPU count, and capital cost for GLM-4.5 or the smaller GLM-4.5-Air, both mixture-of-experts models tuned for agentic and coding workloads.

## Category

on-prem-ai

## Tool

**Kind:** calculator

**Inputs:**

#### GLM-4.5 variant

**Key:** model_total_params_b

**Type:** select

**Default:** 355

**Unit:** B params

**Help:** GLM-4.5-Air trades some capability for a dramatically smaller hardware footprint.

**Options:**

##### GLM-4.5: 355B total, 32B active MoE

**Value:** 355

##### GLM-4.5-Air: 106B total, 12B active MoE

**Value:** 106

#### Weight precision

**Key:** quant_bytes

**Type:** select

**Default:** 1

**Unit:** bytes/param

**Help:** FP8 is the practical production default; INT4 is common for the Air variant on smaller GPU counts.

**Options:**

##### FP16 / BF16 (2 bytes/param)

**Value:** 2

##### FP8 (1 byte/param)

**Value:** 1

##### INT4 / AWQ (0.5 bytes/param)

**Value:** 0.5

#### Serving context length

**Key:** context_tokens

**Type:** slider

**Default:** 32768

**Min:** 4096

**Max:** 131072

**Step:** 4096

**Unit:** tokens

**Help:** GLM-4.5 supports up to 128K context, useful for its agentic and long-document coding use cases.

#### Concurrent active users

**Key:** concurrent_users

**Type:** number

**Default:** 20

**Min:** 1

**Max:** 250

**Step:** 1

**Unit:** users

**Help:** Simultaneous in-flight requests during peak hours.

#### GPU type

**Key:** gpu_vram_gb

**Type:** select

**Default:** 80

**Unit:** GB VRAM

**Help:** VRAM per GPU; enter a matching price below.

**Options:**

##### A100 80GB SXM

**Value:** 80

##### H100 80GB SXM

**Value:** 80

##### H200 141GB SXM

**Value:** 141

##### B200 180GB

**Value:** 180

#### Price per GPU

**Key:** gpu_price

**Type:** number

**Default:** 30000

**Min:** 5000

**Max:** 70000

**Step:** 500

**Unit:** $

**Help:** H100 80GB runs roughly $25,000-$32,000; H200 runs $32,000-$40,000; B200 runs $45,000-$60,000 in 2026.

#### Cloud rental rate per equivalent GPU

**Key:** cloud_hourly_rate

**Type:** number

**Default:** 3.5

**Min:** 0.5

**Max:** 10

**Step:** 0.1

**Unit:** $/hr

**Help:** On-demand H100/H200 rates typically run $2-$6 per hour.

**Outputs:**

#### Model weights footprint

**Key:** model_weights_gb

**Expr:** model_total_params_b * quant_bytes

**Format:** number

**Description:** Full expert set at your chosen precision; active parameters govern compute, total parameters govern memory.

#### KV cache at peak concurrency

**Key:** kv_cache_gb

**Expr:** concurrent_users * context_tokens / 1000 * 0.35

**Format:** number

**Description:** Attention cache growth across concurrent sessions at your target context length.

#### Total VRAM required

**Key:** total_vram_gb

**Expr:** model_weights_gb + kv_cache_gb + model_weights_gb * 0.05

**Format:** number

**Description:** Weights plus KV cache plus a 5% allowance for runtime overhead.

**Primary:** true

#### GPUs required

**Key:** gpus_required

**Expr:** ceil(total_vram_gb / gpu_vram_gb)

**Format:** number

**Description:** Minimum accelerator count for one serving replica at your target concurrency.

#### Hardware capital cost

**Key:** hardware_capex

**Expr:** gpus_required * gpu_price

**Format:** currency

**Description:** One-time GPU spend at your entered price per card.

**Primary:** true

#### Equivalent monthly cloud rental

**Key:** monthly_cloud_cost

**Expr:** gpus_required * cloud_hourly_rate * 730

**Format:** currency

**Description:** What the same GPU count would cost per month rented on-demand, for capex payback comparison.

**Disclaimer:** Planning estimates only. Real memory use depends on serving engine and batch size. Benchmark on representative traffic before finalizing a purchase order.

## Content

**Intro:** This free GLM-4.5 on-prem sizing calculator estimates GPU count, VRAM, and capital cost for Zhipu AI's mixture-of-experts model family, and it is built for teams evaluating GLM-4.5 or its smaller GLM-4.5-Air sibling for agentic and coding-focused deployments. Enter variant, precision, context length, and concurrency, and the tool returns weight footprint, KV cache, total VRAM, GPU count, and cost. GLM-4.5 was tuned specifically for agentic tool-use and coding workloads, and the Air variant exists precisely because many enterprises found the flagship's 355B total parameter footprint more capacity than their agentic workloads actually needed.

**Sections:**

#### How GLM-4.5 runs on hardware you own

**Body:** GLM-4.5 routes 32B active parameters per token through its full 355B total parameter mixture-of-experts architecture, while GLM-4.5-Air routes 12B active through a smaller 106B total. As with any MoE model, the full expert set has to be resident in memory: at FP8, the flagship needs roughly 355GB, requiring 5 H100-class GPUs or 3 H200-class GPUs, while Air needs roughly 106GB, fitting comfortably on 2 GPUs. This meaningful gap between the two variants is why evaluating both against your actual agentic or coding workload before committing hardware matters: many tool-calling and code-generation tasks do not show a large quality gap between them.

**Bullets:**

- GLM-4.5 (355B total) needs roughly 355GB at FP8, requiring 5 H100-class GPUs or 3 H200-class GPUs.
- GLM-4.5-Air (106B total) needs roughly 106GB at FP8, fitting comfortably on 2 GPUs.
- Both variants were tuned specifically for agentic tool-use and coding tasks.
- Active parameters (32B and 12B respectively) govern per-token compute cost and latency.

#### Quantization tradeoffs for GLM-4.5

**Body:** FP8 is the recommended default for both variants on modern datacenter GPUs, offering strong quality retention. INT4/AWQ quantization is particularly attractive for GLM-4.5-Air, since it can bring the smaller variant's footprint down to around 53GB, fitting a single 80GB GPU with room for KV cache, which meaningfully lowers the barrier to entry for agentic AI on-prem. For the full 355B flagship, INT4 quantization reduces the cluster to roughly 3 GPUs, still a real multi-GPU commitment but a substantial reduction from FP8, and worth validating carefully against your specific tool-calling accuracy requirements since agentic workloads are sensitive to subtle output format errors.

**Bullets:**

- FP8 is the recommended default for both variants, with strong quality retention.
- INT4/AWQ brings GLM-4.5-Air to roughly 53GB, fitting a single 80GB GPU.
- INT4 quantization of the flagship reduces cluster size to roughly 3 GPUs.
- Validate quantized checkpoints specifically on tool-calling format accuracy, not just general quality.

#### Serving GLM-4.5 with vLLM and SGLang

**Body:** vLLM and SGLang both support GLM-4.5's architecture with expert-parallel serving for the MoE layers, and given the model's agentic focus, pay particular attention to your serving stack's function-calling and structured-output support, since that is where GLM-4.5 is designed to excel. Agentic workloads also tend to have different traffic patterns than chat: longer sequences of tool calls per user session, more variable request timing, and a higher tolerance for slightly longer per-request latency in exchange for correctness. Size concurrency planning around realistic agentic session patterns rather than assuming chat-style request rates.

**Bullets:**

- vLLM and SGLang both support GLM-4.5's expert-parallel MoE serving.
- Verify strong function-calling and structured-output support in your chosen serving stack.
- Agentic session traffic patterns differ from chat traffic; plan concurrency accordingly.
- Correctness on tool-call formatting often matters more than raw latency for this model's target use cases.

#### When to fine-tune GLM-4.5 instead of prompting

**Body:** Fine-tune GLM-4.5 or Air when your agentic workflows need consistent adherence to a specific tool schema, internal API format, or multi-step task structure that generic prompting handles inconsistently at scale. GLM-4.5-Air is the more practical fine-tuning target for most enterprises given its smaller footprint and faster training iteration, and it is frequently sufficient once fine-tuned on domain-specific agentic examples, such as ERP transaction sequences or structured manufacturing work order flows. Reserve the full flagship for cases where evaluation results show a genuine capability gap on your hardest agentic tasks.

#### How Netray deploys and customizes GLM-4.5 on-prem

**Body:** Netray deploys GLM-4.5 and Air for manufacturers building agentic AI that takes real actions against SyteLine, LN, or M3, where tool-calling correctness directly affects production data integrity. We benchmark both variants against your specific agentic tasks before recommending one, configure expert-parallel serving with function-calling validation, and fine-tune on your ERP transaction patterns where needed. Engagements typically start with a two-week pilot proving reliable tool-calling behavior on a narrow, well-defined workflow before expanding scope.

**Cta:** Get a benchmarked GLM-4.5 deployment plan for your agentic workflows, including a flagship-versus-Air comparison.

**Faq:**

#### Should I default to GLM-4.5 or GLM-4.5-Air?

**Answer:** Start with Air unless you have specific evidence you need the flagship's extra capacity. Air's 106B total parameter footprint is roughly a third of the flagship's, needs far less hardware, and for many enterprise agentic and coding tasks the quality gap is smaller than the hardware cost difference would suggest. Benchmark both against your actual tool-calling workflows before committing to the larger, more expensive deployment.

#### Why is GLM-4.5 specifically good for agentic workloads?

**Answer:** GLM-4.5 was trained and tuned with a particular emphasis on tool-use and multi-step agentic reasoning, meaning its training data and reinforcement learning process specifically optimized for reliable function-calling format, multi-turn tool orchestration, and code generation, rather than being a general-purpose chat model applied to agentic tasks after the fact. This shows up as more consistent structured output formatting, which matters enormously when a model's output is parsed and executed rather than just read by a human.

#### How many GPUs does GLM-4.5-Air realistically need for production?

**Answer:** At FP8, Air's 106B total parameters need roughly 106GB for weights, which fits on 2 GPUs at 80GB each with modest headroom for KV cache and concurrency. At INT4/AWQ, it can fit on a single 80GB GPU. For production workloads with meaningful concurrent user counts, plan for 2 GPUs at FP8 to maintain quality headroom and comfortable KV cache capacity.

#### Does GLM-4.5's Chinese origin create compliance concerns for regulated industries?

**Answer:** Model origin does not by itself create a compliance issue for on-prem deployment, since the weights run entirely inside your network boundary with no data leaving your infrastructure. However, some regulated contracts, particularly in defense, include flow-down clauses restricting the use of technology from specific countries regardless of deployment location. Review your specific contractual and regulatory obligations before adopting any model, and treat this as a legal question rather than a technical one.

## Search terms

GLM-4.5 on-prem, GLM-4.5 GPU requirements, GLM-4.5 hardware calculator, self-hosted GLM-4.5, GLM-4.5-Air deployment, GLM-4.5 quantization, GLM-4.5 fine-tuning, Zhipu AI open model

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json