Quantum-inspired expert pruning
We applied a new quantum-inspired pruning method to the expert layers, reducing the count from 265 to 148 experts per layer.
The method comes out of the tensor-network work that underpins CompactifAI. Rather than scoring experts one at a time and dropping the lowest, it treats expert selection as a joint problem across the layer, so the experts that survive are the ones that matter as a set for the target domain. That distinction matters in MoE pruning, where experts are correlated and greedy per-expert scoring tends to remove redundant capability and keep duplicated capability.
The pruning was targeted: the objective was to retain knowledge relevant to agentic and coding tasks. Capability outside that domain is where the reduction was taken. Quasar 438B is a specialist, and it should be evaluated as one.
Healing
Pruning any model leaves damage, and MoE routing is particularly sensitive to it. After pruning we ran an additional healing pass on the surviving candidate, focused on recovering and sharpening agentic and coding behavior rather than on restoring broad general performance.
This step is what turns a pruned checkpoint into a usable model. Skipping it is the difference between a model that is smaller and a model that is smaller and still good at the job.
Quantization-Aware Compression
We then applied quantization-aware techniques to produce several precision variants of the healed model in FP8 and NVFP4. The current model in the CompactifAI API is in FP8.
Quantization-aware, not post-training: the model is optimized with the target numeric format in the loop, which is why the FP8 build holds up on long agentic traces where accumulated error usually shows first.
What this means in practice
Quasar 438B is a structural change, with a different architecture inherited from GLM-5.2.
What CompactifAI contributes is the pipeline: quantum-inspired expert selection, targeted healing, and quantization-aware compression, applied so that a 265-expert generalist becomes a 148-expert specialist that runs at a fraction of the cost.
Beyond Quasar 438B
Quasar 438B is one output of this pipeline, not the limit of it. CompactifAI is not restricted to public base models. The same compression stages apply to private models: a customer's own model can be compressed and served through the CompactifAI API, with the weights staying under their control.
The pipeline also covers models built from scratch rather than derived from an existing open release, compressed and served the same way. Change the base model, or change the target domain, and the pipeline produces a different model. Quasar 438B is the version we built for coding and agentic work.
More updates are coming for Quasar. Stay tuned for bigger and specialized models.
Get started
CompactifAI API. Sign up and start sending requests: dashboard.compactif.ai
Artificial Analysis. Check out the benchmarks
Contact. Reach out our team at business@multiversecomputing.com