Atlas Inference Engine vs Selvedge

Atlas Inference Engine and Selvedge are both inference engines & infra tracked by AIDiveForge. Below is a side-by-side comparison of pricing, capabilities, platforms, and ownership — sourced from each tool's live website and verified before publishing.

Atlas Inference Engine

The vendor page benchmarks Atlas at 3.1x the decode throughput of vLLM on Nvidia DGX Spark hardware — 111 tok/s average versus 37 tok/s on Qwen3.5-35B, with a cold start measured in two minutes instead of ten. That gap exists because Atlas ships no Python, no PyTorch, and no JIT warm-up: every path from HTTP request to kernel dispatch is compiled. The tradeoff is hardware specificity — hand-tuned CUDA kernels target Blackwell SM120/121, so teams not running DGX Spark get none of the headline numbers. The model matrix covers Qwen, Gemma, Nemotron, Mistral, and MiniMax, but every recipe is written for that hardware profile. Teams running other GPU generations are not the audience.

Selvedge

Selvedge is a local MCP server that AI coding agents (Claude Code, Cursor, Copilot) call as they work, logging the reasoning behind every change into a SQLite file that lives next to your code under .selvedge/. Queries are entity-scoped — you ask about users.email or deps/stripe, not line numbers — so the answer surfaces in the same terms you search in. The vendor describes zero telemetry, no accounts, and no external servers; everything stays on disk. The wall appears when your team needs cross-repo provenance or wants to pipe this data into an existing observability stack — Selvedge emits records but does not integrate with those systems out of the box.

Attribute	Atlas Inference Engine	Selvedge
Pricing	Free	Free
Free trial	No	No
Open source	Yes	Yes
Has API	Yes	No
Self-hosted option	Yes	Yes
Platforms	Linux (Ubuntu 22.04+) with NVIDIA GPU support (Blackwell GB10 primary, Hopper/Ampere in development)	Linux, macOS, Windows (via Python)
Released	—	2026-05
Pros	~2.5 GB container image with no Python or PyTorch dependencies, which means cold starts take two minutes instead of ten — a difference that compounds across every iteration in an agentic development loop. Compiled Rust + CUDA architecture with no GIL or JIT warm-up, so request latency is consistent from the first token rather than degrading during the warm-up window that costs vLLM its first several minutes. Hand-tuned CUDA kernels per model family with NVFP4 and FP8 on Blackwell tensor cores, so quantized inference does not trade throughput for accuracy the way a generic quantization layer would. Multi-Token Prediction speculative decoding built in, so a single DGX Spark node serving a 35B model reaches throughput that would otherwise require additional hardware or a more complex multi-node setup. OpenAI-compatible API endpoint out of the box, so existing tooling — Claude Code, Cline, Open WebUI — connects without a translation layer or custom client code.	Reasoning is captured in the same context window that produced the change — not reconstructed from the diff afterward — which means the intent survives even when the original prompt, the developer who wrote it, and the model version are all gone. Entity-scoped queries (selvedge blame payments.amount, selvedge diff users --since 30d) let you ask about the things you actually search for rather than hunting through line-level history, so a schema audit that would take an afternoon takes a single command. Fully local storage in a SQLite file with no accounts, no telemetry, and no external servers, which means sensitive schema and API change history never leaves the machine — a hard requirement in compliance-heavy environments. Provider-agnostic MCP integration wires into Claude Code, Cursor, and Copilot through a single setup command, so teams already using any of those agents get provenance logging without changing their workflow. Full-text search across all logged events (selvedge search "stripe") and changeset grouping (selvedge changeset add-stripe-billing) mean you can reconstruct the full scope of a feature build after the fact, which is the audit trail that git log alone cannot provide.
Cons	Every published benchmark and kernel optimization targets Nvidia Blackwell SM120/121 on DGX Spark. Teams running Ampere, Ada, or Hopper GPUs get none of the headlined throughput numbers — the architecture constraint is not a tuning issue, it is baked into the kernel design. Those teams are still on vLLM or TensorRT-LLM. The model matrix is a curated, hand-tuned list — Qwen, Gemma, Nemotron, Mistral, MiniMax — not an open registry. A team that needs to serve a fine-tuned model outside that matrix hits a wall immediately and either waits on the Atlas roadmap, opens a Discord request, or returns to vLLM where arbitrary HuggingFace checkpoints load without curation. AGPL-3.0 is the default license. Any team building a closed-source product or operating a SaaS service on top of Atlas is required to obtain a commercial license. Teams that discover this constraint after building on the free version face a licensing conversation before they can ship.	Selvedge has no API and no export integration — teams that need to push reasoning records into an existing compliance platform, a data warehouse, or a centralized observability system must write their own pipeline against the SQLite file, adding a maintenance surface that grows with audit requirements. The store is scoped to a single local project directory; teams running multi-repo codebases where an agent change in one repo depends on a change in another get no cross-repo provenance, and at that point teams managing compliance across repositories will move to a dedicated audit-log solution that operates at the organization level. Selvedge only captures what the agent explicitly logs through the MCP tool call — if an agent skips the log_change call, makes changes outside a supported tool, or the MCP connection drops mid-session, that change has no recorded reasoning and the gap is invisible in the history.

Bottom line

Only Atlas Inference Engine exposes a public API. Choose based on which difference matters most for your workflow.

Frequently asked questions

What is the difference between Atlas Inference Engine and Selvedge?

Atlas Inference Engine is Free and open source, while Selvedge is Free and open source. Compare pricing, free trial, API, platforms, and pros/cons in the table above on AIDiveForge.

Is Atlas Inference Engine better than Selvedge?

It depends on your workflow. Use the side-by-side attributes (pricing, open source, API, self-hosted, platforms) to decide. AIDiveForge does not rank a universal winner — we publish verified facts so you can choose.