Pinned Tweet
Introducing Span-02: anytime reasoning for decisions.
Reasoning models make you pick an effort level and wait. Span-02 takes a deadline and always returns its best decision so far.
It gets better the longer it thinks. With reasoning off, it beats Jev at 1/4 of the price, and runs at 1/3 the latency of decision models at its level.
It can:
- Answer with numbers and scores, not just probabilities and choices
- See images and screens (multimodal)
- Take a configurable reasoning time
And it’s fully open weight!
Respan retweeted
Raced our OSS gateway against Bifrost and LiteLLM on 4 CPU cores.
Requests per second:
Respan Gateway: 61,063
Bifrost: 29,619
LiteLLM: 571
Streaming, the gap gets wider: 35,830 vs 12,999 vs 269.
New in Respan: a SQL editor for your LLM traces.
Every span and trace in your project is now a SQL table: cost, tokens, latency, errors, customers.
Which customers hit the most errors this week?
SELECT customer_identifier,
countIf(status = 'failed') AS errors
FROM logs
WHERE timestamp >= now() - INTERVAL 7 DAY
GROUP BY customer_identifier
ORDER BY errors DESC
- Autocomplete for every table and column
- Chart results and add them to a dashboard
- Export up to 1M rows
Respan retweeted
Shoutout to other open source project!
Respan OSS Gateway: the last AI gateway you ever need.
Because it can connect anything: LLMs, images, video, realtime voice, and MCP tools like web search and long-term memory, all through one endpoint. Each one is just a plugin you switch on.
Against LiteLLM on the same machine:
- 0.08 ms vs 5.9 ms added latency (78x less)
- 61,062 vs 1,348 requests/s with 50 in flight (45x more; LiteLLM on 14 workers)
- 33 MB vs 494 MB memory (15x less; LiteLLM on 1 worker)
Fully written in Rust. MIT/Apache-2.0.
Respan OSS Gateway: the last AI gateway you ever need.
Because it can connect anything: LLMs, images, video, realtime voice, and MCP tools like web search and long-term memory, all through one endpoint. Each one is just a plugin you switch on.
Against LiteLLM on the same machine:
- 0.08 ms vs 5.9 ms added latency (78x less)
- 61,062 vs 1,348 requests/s with 50 in flight (45x more; LiteLLM on 14 workers)
- 33 MB vs 494 MB memory (15x less; LiteLLM on 1 worker)
Fully written in Rust. MIT/Apache-2.0.
Routing and fallbacks are also plugins:
[routes.smart]
to = ["anthropic/claude-haiku-5-5", "openai/gpt-6-luna"]
tries = 2
Claude gets two tries, then it falls back to GPT.
Write a plugin: a guardrail, a smart router, custom billing. It's the same Rust API our own features are built on, and good ones get merged into the main repo.
Try it out, and give us a star if you like it: github.com/respanai/respan-g…
Routing and fallbacks are also plugins:
[routes.smart]
to = ["anthropic/claude-haiku-5-5", "openai/gpt-6-luna"]
tries = 2
Claude gets two tries, then it falls back to GPT.
Write a plugin: a guardrail, a smart router, custom billing. It's the same Rust API our own features are built on, and good ones get merged into the main repo.
Try it out, and give us a star if you like it: github.com/respanai/respan-g…
Routing and fallbacks are also plugins:
[routes.smart]
to = ["anthropic/claude-haiku-5-5", "openai/gpt-6-luna"]
tries = 2
Claude gets two tries, then it falls back to GPT.
Write a plugin: a guardrail, a smart router, custom billing. It's the same Rust API our own features are built on, and good ones get merged into the main repo.
Try it out, and give us a star if you like it: github.com/respanai/respan-g…