Carry the context.
Repeated instructions. Shared documents. Growing conversations. We’re exploring cache reuse to make returning to useful context less expensive.
CONTEXT REUSE OPEN MODELS. FORWARD MOTION.
Inference for
agents that
keep going.
Useful agents do more than answer once.
We’re building open-model inference for long contexts, repeated turns, and the work in between.
In development / Open to design partners
FIG. 01 THE INFERENCE LOOP
Keep the context.
Move the work forward.
Move the work forward.
CONTEXT → COMPUTE → CONTINUATION
THE WAY WE THINK ABOUT INFERENCE
More useful work from every GPU.
01 / THE APPROACH
A better fit for
the way agents work.
One prompt is just the beginning. We look at the whole session—and optimize the system around it.
Keep a steady rhythm.
A fast first token matters. So does everything after it. We evaluate response time, generation speed, and behavior when requests arrive together.
SESSION PERFORMANCETune the whole system.
The model, engine, and hardware should work together. We measure configurations against real workload requirements, then improve what matters.
WORKLOAD-LED OPTIMIZATION02 / LESS REPEATING. MORE DOING.
The next turn
shouldn’t start
from zero.
Agents return to the same instructions and context. An exact shared prefix can be reused, leaving less input to process again.
Try a few turns. See what stays useful.
Reusable prefix New input
agent_sessionCONCEPTUAL EXAMPLE
System instructionsNEW
Repository contextNEW
Current requestNEW
A fresh starting point.
The first request processes its instructions, context, and new input.
An illustration of prefix reuse, not a performance benchmark. Actual reuse depends on the model, serving engine, and exact input prefix.
03 / THE WORK AHEAD
Built around the work.
Not just the request.
We’re interested in teams whose agents do sustained, context-heavy work. These are the workloads shaping our research.
Coding agents
Keep repository context in play through debugging, edits, and the next test run.
Document workflows
Work through substantial source material across extraction, comparison, and synthesis.
Agent systems
Explore and refine over repeated steps, with performance measured across the whole run.
A MEASURED APPROACH
First, understand.
Then, optimize.
Efficiency only counts when it holds up on the workload you actually need to run.
- 01
Map the workload
Context length, session patterns, concurrency, and what a good response time means for your users.
- 02
Establish a baseline
A reproducible configuration and a clear picture of quality, latency, throughput, and resource use.
- 03
Test the improvements
Targeted changes, measured against the same work. Keep the gains that survive the comparison.
EARLY DAYS. AMBITIOUS WORK.
Let’s move
your work forward.
We’re building Odasher and looking for the workloads that will shape it. Start with your model, your context, and where the system slows down.
Create a brief for a future pilot. No account needed.OPEN MODELS. PRACTICAL PROGRESS.IN DEVELOPMENT / 2026
A FEW DETAILS
Before the first turn.
Can I use Odasher today?
Odasher is in development. We’re validating our serving approach before opening a production service. The pilot brief helps define a workload and its success criteria; it does not create an account or reserve GPU capacity.
Which models are you focused on?
Our research focuses on open-weight language models, including the GLM and DeepSeek families. The model and configuration for a future pilot will depend on workload fit, licensing, and deployment support. There is no live model catalog yet.
What will you measure?
Time to first token, generation speed, end-to-end latency, output quality, cache reuse, and resource cost under representative load. We want results that reflect a complete session.
Are performance results or prices available?
Not yet. We’ll share measured results with their workload and configuration once our experiments are complete. Pilot scope and commercial terms will be discussed individually.