> ## Documentation Index > Fetch the complete documentation index at: https://vastai-80aa3a82-auto-openapi-update-b65c39fe.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # Search Offers > Search for available GPU machine offers with advanced filtering and sorting. Each filter parameter (such as `verified`, `gpu_name`, `num_gpus`, etc.) should be an object specifying the operator and value you want to match. **Filter operators:** | Operator | Meaning | Example | |:---------|:-----------------------|:-------------------------------| | `eq` | Equal to | `{ "eq": true }` | | `neq` | Not equal to | `{ "neq": false }` | | `gt` | Greater than | `{ "gt": 0.99 }` | | `lt` | Less than | `{ "lt": 10000 }` | | `gte` | Greater than or equal | `{ "gte": 4 }` | | `lte` | Less than or equal | `{ "lte": 8 }` | | `in` | Value is in a list | `{ "in": ["RTX_3090", "RTX_4090"] }` | | `notin` | Value is not in a list | `{ "notin": ["TW", "SE"] }` | CLI Usage: `vastai search offers 'reliability > 0.99 num_gpus>=4' --order=dph_total` ## OpenAPI ````yaml /api-reference/openapi.yaml post /api/v0/bundles/ openapi: 3.1.0 info: title: Vast.ai API description: >- Vast.ai REST API for managing GPU cloud instances, machine operations, and AI/ML workflows. ## AI Agent Quick-Start Install the CLI skill for your agent (Claude Code, Cursor, Windsurf, etc.): npx skills add vast-ai/vast-cli CLI reference: https://raw.githubusercontent.com/vast-ai/vast-cli/master/vastai/SKILL.md SDK reference: https://raw.githubusercontent.com/vast-ai/vast-cli/master/vastai_sdk/SKILL.md ## Auth All endpoints require `Authorization: Bearer $VAST_API_KEY`. Get your key at: https://cloud.vast.ai/manage-keys/ ## Key Quirks - `gpu_ram` in CLI = GB; in REST API = MB (CLI auto-converts) - SSH keys must be registered BEFORE creating an instance (VM: no recovery; Docker: can add post-create) - `onstart` field is limited to 4048 characters -- gzip+base64 for longer scripts - `POST /api/v0/asks/{id}/` (create instance) returns `new_contract` as the instance ID, not `id` - Poll trap: if `actual_status` becomes `exited`, `unknown`, or `offline` it will never reach `running` -- destroy and retry version: 1.0.0 contact: name: Vast.ai Support url: https://discord.gg/vast servers: - url: https://console.vast.ai description: Production server security: - BearerAuth: [] paths: /api/v0/bundles/: post: tags: - Search summary: Search Offers description: >- Search for available GPU machine offers with advanced filtering and sorting. Each filter parameter (such as `verified`, `gpu_name`, `num_gpus`, etc.) should be an object specifying the operator and value you want to match. **Filter operators:** | Operator | Meaning | Example | |:---------|:-----------------------|:-------------------------------| | `eq` | Equal to | `{ "eq": true }` | | `neq` | Not equal to | `{ "neq": false }` | | `gt` | Greater than | `{ "gt": 0.99 }` | | `lt` | Less than | `{ "lt": 10000 }` | | `gte` | Greater than or equal | `{ "gte": 4 }` | | `lte` | Less than or equal | `{ "lte": 8 }` | | `in` | Value is in a list | `{ "in": ["RTX_3090", "RTX_4090"] }` | | `notin` | Value is not in a list | `{ "notin": ["TW", "SE"] }` | CLI Usage: `vastai search offers 'reliability > 0.99 num_gpus>=4' --order=dph_total` requestBody: content: application/json: schema: type: object properties: limit: type: integer description: Max offers to return type: type: string enum: - ondemand - bid - reserved description: > Instance type for the offer. Affects pricing calculation. - **ondemand**: Fixed pricing based on listed rates. Default. - **bid** (interruptible): Uses minimum bid price. Lower cost but may be interrupted if outbid. - **reserved**: Reserved instance pricing. verified: type: object description: Machine verification status properties: eq: type: boolean default: eq: true rentable: type: object description: Whether machine is rentable properties: eq: type: boolean default: eq: true rented: type: object description: > When set to true, include offers where the calling user already has rented GPUs. This is useful for finding offers on machines you're already renting. properties: eq: type: boolean default: eq: false gpu_name: type: object description: >- GPU model name. Example: {"eq": "RTX_4090"} or {"in": ["RTX_3090", "RTX_4090"]} properties: in: type: array items: type: string example: - RTX_3090 - RTX_4090 reliability: type: object description: 'Machine reliability score (0-1). Example: {"gte": 0.99}' num_gpus: type: object description: 'Number of GPUs. Example: {"gte": 4} or {"in": [1, 2, 4, 8]}' properties: in: type: array items: type: integer example: - 1 - 2 - 4 - 8 gpu_ram: type: object description: 'GPU RAM in MB. Example: {"gte": 24000}' duration: type: object description: >- Minimum required rental duration in seconds (the offer must be available for at least this long from now). machine_id: type: object description: Filter by specific host machine ID dlperf_per_dphtotal: type: object description: DLPerf per dollar per hour dph_total: type: object description: 'Total $/hour rental cost. Example: {"lte": 0.5}' flops_per_dphtotal: type: object description: TFLOPs per $/hour geolocation: type: object description: >- Machine location (two letter country code). Example: {"in": ["US", "CA"]} properties: in: type: array items: type: string example: - US - CA gpu_arch: type: object description: >- Host machine GPU architecture (e.g. nvidia, amd). Example: {"eq": "nvidia"} dlperf: type: object description: Deep Learning performance score cuda_max_good: type: object description: Maximum supported CUDA version inet_down: type: object description: Download bandwidth (MB/s) inet_up: type: object description: Upload bandwidth (MB/s) inet_down_cost: type: object description: Download bandwidth cost ($/GB) inet_up_cost: type: object description: Upload bandwidth cost ($/GB) driver_version: type: object description: NVIDIA driver version in the format "XXX.XX.XX" compute_cap: type: object description: >- CUDA compute capability x 100. Use 650 for compute capability 6.5 or 700 for 7.0. cpu_arch: type: object description: >- Host machine CPU architecture (Default is amd64). Example: {"eq": "amd64"} has_avx: type: object description: CPU supports AVX instruction set cpu_cores: type: object description: Number of virtual CPUs cpu_cores_effective: type: object description: Effective vCPU count for the offer cpu_ghz: type: object description: CPU clock speed in GHz cpu_ram: type: object description: CPU RAM in MB datacenter: type: object description: Show only datacenter offers external: type: object description: Show external offers in addition to datacenter offers disk_bw: type: object description: Disk read bandwidth in MB/s disk_space: type: object description: Disk storage space in GB bw_nvlink: type: object description: >- NVLink interconnect bandwidth in GB/s. Use when filtering for multi-GPU systems with high-speed NVLink. gpu_max_power: type: object description: GPU power limit in watts gpu_max_temp: type: object description: GPU temperature limit in Celsius gpu_mem_bw: type: object description: GPU memory bandwidth in GB/s gpu_total_ram: type: object description: Total GPU RAM across all GPUs in MB gpu_frac: type: object description: Fraction of the total GPU resources being offered gpu_display_active: type: object description: Whether the GPU has an attached display direct_port_count: type: object description: Number of direct ports host_id: type: object description: Host user ID id: type: object description: Offer ID min_bid: type: object description: Minimum bid price ($/hour) mobo_name: type: object description: Motherboard name pci_gen: type: object description: PCIe generation pcie_bw: type: object description: PCIe bandwidth (CPU to GPU) storage_cost: type: object description: Storage cost in $/GB/month static_ip: type: object description: Whether the IP address is static/stable total_flops: type: object description: >- Total theoretical GPU compute performance (TFLOPs) across all GPUs. os_version: type: object description: Host machine Ubuntu OS version ubuntu_version: type: object description: Host machine Ubuntu OS version (alias for os_version) verification: type: object description: >- Machine verification status string (verified, deverified, unverified). Example: {"eq": "verified"} vms_enabled: type: object description: Whether the machine is a VM instance order: type: array description: | List of sort fields and directions. Each entry should contain two elements: 1. The field name to sort by (string) 2. The sort direction ("asc" or "desc") items: type: array items: type: string minItems: 2 maxItems: 2 allocated_storage: type: number description: > Storage allocation size in GB for the instance. This sets the disk size when creating the instance and cannot be changed later. Default is 8GB. disable_bundling: type: boolean description: If true, machines are not bundled into multi-GPU offers. additionalProperties: type: object description: >- Filter expression for an offer field: {operator: value}, e.g. {"gte": 2} or {"in": ["RTX 4090", "RTX 5090"]}. properties: eq: description: '"eq" comparison against the field value' neq: description: '"neq" comparison against the field value' gt: description: '"gt" comparison against the field value' gte: description: '"gte" comparison against the field value' lt: description: '"lt" comparison against the field value' lte: description: '"lte" comparison against the field value' in: description: '"in" comparison against the field value' notin: description: '"notin" comparison against the field value' additionalProperties: false example: limit: 100 type: on-demand verified: eq: true rentable: eq: true rented: eq: false description: The raw search query object (not wrapped in a "q" key). responses: '200': description: 'Successful response with matching offers: {"offers": [...]}' content: application/json: schema: type: object properties: offers: type: array description: >- Array of offer objects with pricing, GPU specs, and scoring fields items: type: object '400': description: >- Invalid query or body: "Invalid query: must be dict of conditions" or "Invalid json body" content: application/json: schema: $ref: '#/components/schemas/Error' security: - BearerAuth: [] components: schemas: Error: type: object properties: error: type: string msg: type: string securitySchemes: BearerAuth: type: http scheme: bearer description: API key must be provided in the Authorization header ````