Engineering ready
GLM-4.6V FP8 Production Package
GLM / GLM-4.6V · fp8 · Engineering status: Engineering ready
Transformation lineage
Transparent pricing
Prices are shown in your selected display currency. This configuration is quoted in USD; the accepted quote preserves its currency, prices and terms.
- Base model engineeringUSD 59,160
- FP8 QuantizationUSD 5,900
- Runtime packagingUSD 2,800
- Evaluation and benchmarkingUSD 3,600
- Download packageUSD 1,500
- Managed deployment setupUSD 3,200
- One-time engineeringUSD 76,160
Monthly hosting
- 2 x NVIDIA H100 SXM 80GBUSD 3,720
- Monthly hostingUSD 3,720
Engineering scope and evidence
Orderable as scoped engineering
You may configure and request this engineering engagement. Source verification, final deployment suitability and acceptance criteria are confirmed through the existing scope-review process before implementation.
Source release
- Model family
- GLM
- Base release
- GLM-4.6V
- Package type
- Quantized
Source checkpoint precision
Defined during the agreed engineering work
The release has more than one registered checkpoint format. The exact input checkpoint and revision must be agreed during scope validation before engineering starts.
Target precision
Defined in the catalogue
The package specification defines FP8 as the target format. This is an engineering specification, not evidence of a completed artifact or measured quality.
Deliverable scope
Defined during the agreed engineering work
This purchase commissions the listed engineering and evaluation work for the exact release. The selected delivery mode and accepted quotation define the handover; it is not an immediate download of an already completed artifact.
- Validated transformed model artifact
- Tokenizer and model configuration
- Runtime configuration for the selected delivery
- Checksums and version manifest
- Benchmark report
Download: delivery of the agreed model artifacts, tokenizer/configuration, version manifest and checksums after engineering and validation, subject to the source licence.
Host: deployment of the agreed result on the selected hosting configuration. GPU allocation and recurring charges are quoted separately from one-time engineering.
Download + Host: the agreed artifact handover and hosted deployment, with their respective licence and hosting requirements.
Evaluation: the configured evaluation work and its resulting report. No performance result or quality-retention figure is promised before measurement.
Target GPU
Defined in the catalogue
The configuration targets 2 × NVIDIA H100 SXM 80GB. A reference specification or estimated fit does not reserve rental stock; inventory and deployment suitability are checked before allocation.
Memory fit
Defined during the agreed engineering work
These are stored sizing estimates for the selected GPU count, context and workload, not measured deployment results. Runtime overhead and real workload behaviour require validation.
Estimated model weights: 105.33 GB
Estimated KV cache: 2 GB
Estimated required VRAM: 141.43 GB
Estimated available VRAM: 160 GB
Benchmark evidence
Defined during the agreed engineering work
No complete measured benchmark is recorded for this package. Measurements are produced during the agreed evaluation using the exact result, hardware, runtime and settings. Estimates are not measured results.
Delivery mode
Defined in the catalogue
Download + host
What the engineering status means
Engineering ready: the configured engineering workflow and commercial components can be scoped. It does not mean the resulting artifact has already been built or benchmarked.
Registered checkpoint evidence
- glm-4-6v-bf16 · BF16 · Registry verification date: 2026-09-03
Source revision
4e2d47eb0b41c5280d8294b17cef9e94fdcfff46 - glm-4-6v-fp8 · FP8 · Registry verification date: 2026-08-26
Source revision
33172e26eb88482cf3d0a36fced01d05454734ec
Delivery limitations
Configured delivery: Download + host
Delivery options: Download package / Host with FusionPointAI / Download + host
Transformation lineage
Source precision: Defined during the agreed engineering work
Target precision: FP8
Transformation operations: FP8 Quantization, Runtime packaging, Evaluation and benchmarking
GPU topology
Target GPU: NVIDIA H100 SXM 80GB
GPU count: 2
Runtime packaging: transformers