Canada

LightspeedLM

Infinite context, bounded cost.

A long-running agent should keep the thread of its own work: the corrections it was given, the reasoning behind decisions it made. LightspeedLM, from the research team at NexSpring, holds that continuity at the inference layer. The active footprint stays bounded as history grows, and nothing is summarized away or replayed to get it back: the agent works from the state it produced, not a paraphrase of it.

Ask us about the upcoming private beta.

1.5B native limit 32k
With LightspeedLM 448k
Current record 14×
How far back a fact can sit and still be recalled. Early result on a local 1.5B model with a 32k window: facts were found as far as 448k tokens back.

The continuity gap

Agents can search documents, retrieve facts, and summarize past conversations. A long-running one still loses the thread of its own work, because every way of handling a growing session forces a trade.

Keep history active
Memory, latency, and inference work grow with it.
Replay old material
The agent pays again to rebuild a state it already reached.
Summarize or compact
Details, rationale, and corrections go missing.
Start a new session
It comes back with records of the past and none of the continuity.

The result is familiar: repeated mistakes, lost decisions, stale instructions, and people rebuilding context before work can resume. Retrieval and explicit memory are still essential for searchable facts, and long-context models keep more within reach. Neither closes the continuity gap.

Continuity at the inference layer

What the model can see at any moment stays the same size however long the session runs. That is the active buffer, holding the key and value vectors the model has already computed for the tokens it is working from. Everything else moves to long-term storage, and KV splicing moves entries between the two in place, without re-running the model over them. Three moves do the work:

Keep
The cache persists between turns. A new turn appends only its new tokens, and nothing already computed is run again.
Trim
At the budget limit the engine evicts the least-attended entries rather than the oldest, and offloads them instead of discarding them.
Re-pack
Retained entries are re-positioned to stay inside the trained window, so a session can run past the model's native limit.
Activein the cache
Long-termcheap, deep
Cold entries move out to long-term storage and are spliced back when a step needs them, so nothing is dropped to make room.

Drops into your harness

The private beta ships with an OpenAI-compatible API, so the agent framework you already run works against it without changes. Point your client at it and keep your prompts, tools, and orchestration as they are. It runs self-hosted or in your own cloud, so session state stays inside your deployment.

What it changes

More consistent work
Corrections, conventions, and task history carry across long workflows, so there are fewer repeated mistakes and less drift from decisions already made.
Less replay
Less repeated processing, and less time spent restoring context the agent already had.
Fits your memory stack
It sits alongside retrieval, files, and application memory. Searchable facts stay explicit; continuity covers what those systems miss.

Who it is for

Agent developers

Sessions measured in days.

Companions and coding agents that run for hours or days, keeping the character and decisions they built up.

Platform teams

Serving cost that stays flat.

Only new tokens are prefilled and the active buffer is fixed, so per-step work does not scale with lifetime history. Throughput against session length is what the beta will measure.

Status

LightspeedLM runs today as an experimental research prototype, and we are putting together a private beta for a small number of teams. It runs on a CUDA GPU against a Qwen2 or Qwen3 checkpoint. Tell us the framework you drive inference from and the workload you want to keep alive, and we will scope the beta around it.

On HotpotQA, in the distractor setting, an 8k active buffer scored 0.70 token-F1 against 0.65 for the same questions run with the full context. Eviction on its own, with nothing recalled after it left the buffer, scored 0.28. These are preliminary results on Qwen3-32B over 20 questions, and they have not been independently validated.

It is our first step toward agents that carry their experience forward without giving up speed or control.

Ask us about the upcoming private beta, and tell us what you are running and how long it needs to last.