☄️ Inside Quasar 438B
Quasar 438B is a compressed and specialized model built from GLM-5.2, the open-weights model developed by Z.ai.
Using CompactifAI, we reduced each MoE layer from 265 to 148 experts through quantum-inspired joint expert selection.
We ealing to recover and sharpen coding and agentic behavior, followed by quantization-aware compression to produce FP8 and NVFP4 variants.
📰 Read the technical note: multiversecomputing.com/papers/inside-…
📍Access the model: auth.multiverseapp.ai/sign-up?redire…
Quasar 438B is a compressed and specialized model built from GLM-5.2, the open-weights model developed by Z.ai.
Using CompactifAI, we reduced each MoE layer from 265 to 148 experts through quantum-inspired joint expert selection.
We ealing to recover and sharpen coding and agentic behavior, followed by quantization-aware compression to produce FP8 and NVFP4 variants.
📰 Read the technical note: multiversecomputing.com/papers/inside-…
📍Access the model: auth.multiverseapp.ai/sign-up?redire…
Community note
Multiverse Computing initially claimed this was a "self-developed", "leading european" model.
After the backlash in x.com/MultiverseComp… they now admit it is in fact a GLM 5.2 REAP.
After the backlash in x.com/MultiverseComp… they now admit it is in fact a GLM 5.2 REAP.
Sep 3, 2026 · 2:23 PM UTC
41
12
98
75,232
@MultiverseCompu this is why cost per task should be the metric
0
0
0
247
@MultiverseCompu If smaller and more efficient, why more expensive than GLM 5.2? 🤔
3
0
33
2,695
@MultiverseCompu You literally took GLM-5.2, Made it WAY worse, (look at Tokens per Task and Cost per task being literally fucking red while scoring worse than all models) and you still post about it
This is embarrasing
This is embarrasing
1
0
30
1,805
@MultiverseCompu Had to clarify after all the backslash. Shameful.
0
0
27
1,366
@MultiverseCompu Calling it “the most advanced European model” and then it's a compressed version of GLM-5.2 sounds like a joke
0
0
24
1,045
@MultiverseCompu Sois una vergüenza. De "el mejor modelo europeo" a ser una simple cuantizacion de glm 5.2. Que tiene su mérito, pero mentisteis.
1
1
20
1,125
@MultiverseCompu Why does this deserves a specific name? It is GLM-5.2-438B, not a new model
0
0
15
1,358
@MultiverseCompu Inadmissible !
Dans ce cas, vous n’avez pas le droit de dire que c’est un modèle européen… vous avez menti
Dans ce cas, vous n’avez pas le droit de dire que c’est un modèle européen… vous avez menti
0
0
6
466
@MultiverseCompu Lo peor es que habéis estado vendiendo humo con tecnologías "inspiradas en la cuántica" que en realidad son técnicas ampliamente conocidas que cualquiera puede hacer con un hardware razonable y sin necesidad de millones de euros de financiación pública. Bochornoso.
0
1
4
252
@MultiverseCompu Even more embarassing that the original announcement, I bet no one can find a single valid reason to use this variant at this price rather than the GLM-5.2 hosted by Mistral (docs.mistral.ai/models/zai-glm…), or literally any other model if you're not bound by strict European laws
1
0
2
724
@MultiverseCompu So GLM 5.2, but significantly worse and significantly more expensive. Wtf?
0
0
1
512
@MultiverseCompu You claim no loss in quality but the benchmarks are way off from GLM 5.2
0
0
1
1,469
@MultiverseCompu We would love to run this model on our
Benchmark.
Benchmark.
0
0
0
362
@MultiverseCompu la falta de transparencia de multiverse computing es interesante. en el lanzamiento no se mencionaba GLM-5.2. hasta que se les echaron encima.
cito:
"Introducing Quasar 438B: Europe's Leading AI Model"
"Quasar 438B is our flagship reasoning model"
multiversecomputing.com/resources/intr…
cito:
"Introducing Quasar 438B: Europe's Leading AI Model"
"Quasar 438B is our flagship reasoning model"
multiversecomputing.com/resources/intr…
0
0
0
52
@MultiverseCompu Knowing that Qwen3.8:27b is smarter is crazy...
0
0
0
625
@MultiverseCompu Sois una decepción y vergüenza para la IA española y Europea. Os habéis convertido en el hazmerreír del sector y habéis dañado la imagen del ecosistema hispano con vuestro burdo intento de engaño.
100M más del gobierno para vosotros para intentarlo de nuevo venga...
100M más del gobierno para vosotros para intentarlo de nuevo venga...
0
0
10
472
@MultiverseCompu "quantization aware compression" lmao you wanted to write "quantization" but it was too short to sound cool
0
0
5
393
@MultiverseCompu Sois el perfecto ejemplo de innovacion europea, copiar un modelo, empeorarlo y todo con dinero publico.
Basura es lo que sois.
Basura es lo que sois.
0
0
6
448
@MultiverseCompu u are funkin late and using glm 5.2 as base??? LMAOOOOOOOOOOOOOOOOOOOOOO
0
0
2
271
@MultiverseCompu This is utterly ridiculous. Probably grabbed a dozen public funds to end up announcing the "sovereign quantized asian model of Europe".
0
0
3
353
@MultiverseCompu Lo que no entiendo es por que no dijisteis esto ayer.
0
0
3
596
@MultiverseCompu Lo que habéis hecho os lo hago yo en un fin de semana con tan solo una jarra de café y durmiendo muy poco. Eso sí, os lo habéis llevado bien fresco. A por la Series A, campeones.
Y pedidle financiación al Santander, que os la darán fijo.
Y pedidle financiación al Santander, que os la darán fijo.
0
0
2
185
@MultiverseCompu Curious how router entropy held up after dropping ~44% of the experts. Did you need much fine-tuning to prevent collapse on niche tokens?
1
0
0
522
