Supported Models

Elastic is a curated catalog of production models. Dedicated runs any Hugging Face model you choose. Use display names (e.g. claude-sonnet-5, kimi-k3, grok-4.3) in API requests.

Elastic Inference Catalog

Hypervize Elastic Inference offers a curated, opinionated catalog of production models from leading providers. We select and maintain models that deliver reliable performance, strong tool-calling behavior, and excellent real-world results.

(For training and large-scale workloads see our bare metal and compute offerings. For any Hugging Face model you want, use Dedicated endpoints.)

Key Providers & Highlights

• Moonshot — Kimi K3 (flagship; 1M context; long-horizon coding + knowledge work; vision + tools; Chat Pro tier)
• Anthropic — Claude Fable 5, Sonnet 5, Opus 5, Haiku 4.5 (vision + text)
• Meta — Llama 4 Scout, Llama 3.3 70B, Llama 3.2 11B (vision)
• Mistral — Mistral Large 3, Devstral, Ministral
• DeepSeek — DeepSeek V3.2 (catalog; available under All models)
• OpenAI — GPT-5.6 Sol, GPT-5.5, GPT-5.5 Pro (direct OpenAI API; frontier reasoning + vision)
• xAI — Grok 4.5, Grok 4.3, Grok Reasoning
• Qwen — Qwen3 Coder, Qwen3 VL, Qwen3 Next, Qwen3 32B
• Cohere — Embed models (text embeddings via /api/embeddings)
• NVIDIA — Nemotron models
• Google — Gemma 4 31B (Chat budget pick), Gemma 3 family
• AI21, MiniMax, OpenAI OSS, Amazon Nova, Stability AI, TwelveLabs, Writer — Full coverage

Kimi K3 (Moonshot)

Use display name kimi-k3 or full value moonshot.kimi-k3 in /api/chat/completions.

How to Reference Models

Recommended: Use the display name (e.g., claude-sonnet-5, qwen3-32b, grok-4.3).

Hypervize maps display names to the correct model automatically. You may also pass a full catalog identifier (e.g. xai.grok-4.3) if you have a specific reason.

Fallbacks under provider rate limiting

On Elastic, some models may automatically fall back to another catalog model if the primary provider is rate-limited or temporarily unavailable. This keeps requests succeeding when possible; it is not available for every model. When a fallback runs, the response model field and usage metering follow the model that was served — see Elastic Inference and Billing & Payments.

Pricing & Context

Every model entry includes:

• Input and output cost per million tokens (or per image for vision models)
• Recommended max_tokens (include this, especially for tool/agent loops — omitting it can lead to very long generations)
• Context window size (where known)

These values appear in the dashboard model picker and feed estimated cost in the UI.

Embedding models (e.g. Cohere cohere.embed-v4) only consume input tokens and are billed on prompt_tokens / total_tokens from the embeddings response. Use them via the standard /api/embeddings endpoint (OpenAI compatible).

Dedicated Endpoints — Any Hugging Face Model

When you deploy a Dedicated endpoint, you get full flexibility — this is where you access the full breadth of models beyond our curated Elastic catalog. Hypervize analyzes the model (Hugging Face metadata and known recipes where available) and configures serving automatically:

• Text LLMs: expert parallelism for MoE, speculative decoding for latency, tool/reasoning parsers, prefix caching and chunked prefill for long context, and related settings.
• Image generation: Diffusers text-to-image pipelines on dedicated GPU capacity; OpenAI-shaped POST /api/d/{id}/images/generations. Weekly GPU-hour billing only (no prepaid per-image charges).

You can deploy:

• Any public Hugging Face text model (e.g., meta-llama/Meta-Llama-3-8B-Instruct)
• Diffusers text-to-image models (see matrix below)
• Gated models (by supplying a Hugging Face access token at deploy time)
• Your own fine-tuned or merged models (upload to HF or use private storage in future)

Elastic is the curated catalog. Dedicated text runs any Hugging Face chat LLM with automatic tuning. Dedicated image supports Diffusers text-to-image pipelines, including common LoRA-on-base setups when a base model can be resolved.

Hardware is automatically sized based on:

1. Text: model weight size, name-based heuristics, and known serving recipes
2. Image: family VRAM needs and GPU count tiers (not inflated Hub storage totals)

Current tiers (subject to change):

• 1× / 2× / 4× / 8× A10G-class GPUs
• 8× A100
• 8× H100
• 8× B200 (Blackwell)

Dedicated image model matrix

Default steps, guidance, and GPU sizing are chosen from the model family so you get sensible defaults without hand-tuning the container.

Scale-to-zero: off by default for image endpoints; cold starts can take minutes while weights reload. Prefer always-on for interactive use.

Vision & Multimodal

Several Elastic models support images. Send them using the standard OpenAI vision message format:

{
  "role": "user",
  "content": [
    { "type": "text", "text": "Describe this image" },
    { "type": "image_url", "image_url": { "url": "https://..." } }
  ]
}

Dedicated endpoints support vision if the underlying model and our inference runtime configuration support it.

Model Updates & Deprecations

• The Elastic catalog is updated periodically as we evaluate and add new high-quality models that meet our standards.
• When a model is deprecated or retired we will announce a migration window with recommended alternatives from our curated set.
• Dedicated endpoints are yours — you control the model version completely.

Requesting New Models (Elastic)

If you need a specific model that is not yet available in the Elastic catalog, open a support request or enterprise conversation. We evaluate and add high-demand models on a regular cadence.

Related

• Elastic Inference API
• Dedicated Inference Endpoints
• Provisioning Dedicated Endpoints