Superseded
Mercury Coder was Inception Labs’ first coding diffusion model. Inception’s current line (docs and blog, 2026-10-02) is Mercury 2.5 (2026-09-08), Mercury 2 (2026-02-24), Mercury Edit 2 (next-edit/coding, 2026-03-30) and Mercury Voice (2026-09-29). The docs no longer mention Mercury Coder.
Mercury Coder (and the Mercury family)
by Inception Labs (no company note in the vault)
Diffusion-based language models: instead of generating tokens one by one, they refine a whole answer in parallel, which the vendor says gives high throughput.
Current Mercury line (Inception blog and docs, checked 2026-10-02)
- Mercury 2 (2026-02-24): “the world’s fastest reasoning language model”; later posts cover Azure AI Foundry availability (2026-06-24), sub-300 ms time-to-first-token on standard NVIDIA GPUs (2026-07-14) and a search use case (2026-08-11).
- Mercury Edit 2 (2026-03-30): next-edit diffusion model for coding workflows. The API has chat, fill-in-the-middle (FIM) and code-edit endpoints.
- Mercury 2.5 (2026-09-08): 1,107 tokens/s on NVIDIA GPUs, 260K context, tunable reasoning, parallel tool calls and JSON-schema outputs; priced per token (usage-based, with a launch promotion); positioned by Inception against cost-optimised models such as GPT-5.6 Luna and Claude Haiku 4.5 (vendor claim). Available through Inception’s API, Baseten and OpenRouter; OpenAI-compatible API; integrations include Cursor, Zed and LangChain.
- Mercury Voice (2026-09-29): real-time reasoning model for voice agents (under 170 ms time-to-first-token per the Mercury 2.5 post).
- Funding: $50M seed round announced 2025-11-06 (Inception blog).
Mercury Coder (historical)
The original Mercury Coder was Inception’s code-generation diffusion model, with a free chat demo at chat.inceptionlabs.ai. The older figures in this note (700-1,100 tokens/s, 32K context, older pricing) described that 2025 model and are not current; nothing in the 2026 docs lets them be re-verified, so they were removed.
Why diffusion matters
Inception positions parallel, iterative refinement as suited to IDE completions, next-edit suggestions, real-time agents and voice (vendor claim). In the author’s view (opinion), a trade-off is a newer architecture with a smaller tooling ecosystem than autoregressive models.
Related
Coding tools that integrate Mercury (Cursor, Zed) are covered in their own notes.
Open items
- Whether the
mercury-coderAPI model is still served: re-checked 2026-10-05, docs.inceptionlabs.ai/llms.txt does not mentionmercury-coderand gives no deprecation notice; model IDs are not enumerated there, so still unverified. - Benchmarks and speed claims in this note are vendor statements.