GPU · active compute
Keeps active experts and the shared model computation near the processor. The GPU is the fast working set, not the entire model warehouse.
Algodyne Corporation / Intelligence infrastructure
ASTRA.
Large-model intelligence. Local execution. A path to models that learn.
Astra runs large open-weight models on hardware you control. It coordinates GPU, CPU, RAM, and SSD to make 125-billion-parameter-class inference practical on a workstation. Liquid Weights is the next research frontier: selectively adapting actual model parameters from verified experience.
Fig. 01 / Astra runtimeMeasured locally
Intelligence for a brighter tomorrow01GPU / compute
02RAM / experts
03SSD / sparse data
01 / Model class125BQwen3.8-Flash-Next, IQ3_XXS quantization.
02 / GPU memory24 GBRTX 5090 Laptop with approximately 125 GB host RAM.
03 / Output tokens/s74–86Measured 128K-context essay and code generation workloads.
04 / Context tested2M2,048K setting: 4/4 retrieval and 1/1 aggregation tests passed.
01 / Execution
One model. Three memory tiers. No datacenter required.
Astra's runtime keeps the fast path on the GPU while using system memory and sparse storage for the much larger model footprint.
RAM · expert residency
Holds the model’s expert weights in system memory. CPU execution handles experts outside the GPU working set concurrently.
SSD · sparse storage
Serves the large lookup table through sparse reads rather than requiring it to live entirely in accelerator memory.
Measured boundaries
The numbers above are internal results for a particular quantized Qwen model and a workstation with approximately 125 GB of RAM, not a claim that 125B parameters fit entirely in 24 GB VRAM. At the 2M context setting, retrieval passed 4/4 and aggregation passed 1/1; a three-hop reasoning test failed at both short and long contexts. Initial 2M prefill took roughly 32 minutes. Throughput figures above were measured at 128K.
02 / Research frontierExperimental · not shipped
Liquid Weights. Intelligence that can change without losing its foundations.
Today's deployed models usually keep their trained parameters fixed. Astra Liquid Weights investigates persistent, reversible parameter updates selected from independently verified knowledge and evaluated before promotion. This is numerical model adaptation—not a retrieval index, a prompt shortcut, or a claim of unrestricted autonomous learning.
Select
Identify a stable, attributable claim or rule. Unsupported output never becomes training truth.
Adapt
Train an isolated numerical parameter overlay without mutating the original model checkpoint.
Challenge
Test unseen paraphrases, counterexamples, interference, forgetting, and unrelated capability.
Retain or revert
Persist accepted edits across restart. Disable, re-enable, or reject them independently.
What must be demonstrated
A genuinely new fact learned into numerical parameters must survive process restart without retrieval or prompt injection; isolated edits must be disableable and recoverable. The experiments also measure forgetting, leakage into nearby answers, generalization, and runtime cost. Until that evidence exists, Liquid Weights remains a research program.
03 / Applications
For systems that need to think on their own terms.
From local development to privately operated enterprise intelligence, Astra is designed around control over the runtime and the evidence behind its performance.
Private inference
Run capable open-weight intelligence locally where data control and operational independence matter.
Long-context analysis
Work across large codebases and document collections with an evidence-backed context envelope.
Coding and agents
Serve local applications and coding agents through existing OpenAI- and Anthropic-compatible interfaces.
Adaptive specialists · research
Explore whether verified operational experience can become durable, reversible model capability.
04 / Work with Algodyne
Start a conversationBuild on intelligence you control.
Interested in high-performance private inference, commercial deployments, technical validation, or backing the research? Talk directly with Algodyne Corporation. Benchmark evidence and experimental findings can be shared with qualified technical collaborators.
ASTRA / ALGODYNE CORPORATION — BUILD SYSTEMS. LAUNCH VENTURES. COMPOUND IMPACT.