Ling-3.0-flash-Sante: Research LLM, Setup and Memory

Sante produces text, supports reasoning and can use tools through compatible serving frameworks. Its hybrid architecture combines Kimi Delta Attention, gated MLA and sparse MoE layers. Its training and evaluations specialize in healthcare research.

The model card specifies 124B total parameters, 5.1B active per token and training to a 256K context. The configuration sets max_position_embeddings to 262144 and uses BF16. Active parameters describe computation for a token, not how many weights must be stored. The public repository occupies about 255GB; allow additional runtime and request memory when provisioning a server.

The SGLang quickstart uses lmsysorg/sglang:dev-Ling-3.0-flash with tensor parallelism across four 141GB-class GPUs or a compatible four-GPU Blackwell node. For 80GB H100 or H800 GPUs, the card specifies eight-way tensor parallelism. These are documented example configurations, not a measured minimum or a promise that every context and concurrency setting fits.

The SGLang example serves 262,144 tokens, uses NEXTN speculative decoding and reserves a static memory fraction of 0.8. The vLLM path requires a build with Ling-3.0 support and uses ling3 for both tool-call and reasoning parsing, with MTP speculative decoding. Record the installed build or image digest: development tags change over time.

Start the server on infrastructure you control, keep the endpoint private and confirm a short request before increasing context. Thinking is enabled by the chat template by default. The card allows chat_template_kwargs.enable_thinking=false for a request that does not need reasoning. Check that your client preserves the chosen mode and parses tool calls before adding a research tool.

Local operation includes all allocated GPUs while weights load, prompts are processed and research tools run. Add storage and external tool charges, then divide by accepted tasks. A long-context or multi-agent run can spend substantial time outside token generation; a decode rate alone cannot price the workflow.

For an illustrative allocation costing $12/hour in total, 20 minutes costs $4 before storage and external tools. Two accepted document comparisons from that session cost $2 each. Keep the total hourly rate of the node separate from the rate of one GPU.

Use a small reference set with known source passages. Track unsupported statements, extraction errors, tool failures and the time spent checking the output. Compare against the general Ling model on the same documents if specialization is the reason for choosing Sante.

The developer reports healthcare, reasoning, deep-research and tool-use evaluations. Most use temperature 1.0 and top_p 0.95 with the highest reasoning tier. The BrowseComp multi-agent evaluation uses an in-house harness with up to 64 sub-agents per task; that result includes more than the standalone model.

The linked community announcement discusses the specialization and benchmark comparisons, but provides no reproducible deployment memory or per-task cost measurements. The model card identifies Sante as a research model, not a medical device; use it for research and source checking rather than independent diagnosis, prescribing or treatment decisions.