Model library / sabicap
Sabi-Noema-MLA is a 27.4 B-parameter agentic model derived from Qwen3.8-27B, using multi-head latent attention to reduce attention-cache memory.
Parameters in total
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
Sabi-Noema-MLA, released by sabicap, is described as an agentic model built on Qwen3.8-27B and converted to multi-head latent attention (MLA). The conversion shrinks the attention cache to 18 KiB per token (9 KiB with simulated FP8), allowing longer reasoning sessions to run on consumer-grade GPUs.
The model has 27.4 billion parameters, a context window of 262,144 tokens, and is licensed under Apache-2.0. It contains no expert layers and its weight format is not published. With 4-bit weights it can host seven concurrent 32 K-token sessions on a 24 GB GPU, and the attention cache size is 18 KiB per token (9 KiB with FP8).
Specification
| Repository | sabicap/sabi-noema-mla |
| Publisher | sabicap |
| Published | 13 October 2026 |
| Architecture | SabiMlaForConditionalGeneration |
| Model type | sabi_mla |
| Parameters | 27,356,728,560 |
| Layers | 64 |
| Hidden size | 5,120 |
| Attention heads | 24, 4 key/value heads |
| Context length | 262,144 tokens |
| Vocabulary | 248,320 tokens |
| Licence | apache-2.0 |
| Base model | sabicap/sabi-noema |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 51.0 GiB | Full precision as released |
| fp8 | 25.5 GiB | 8-bit, near-lossless on most models |
| int4 | 12.7 GiB | 4-bit, smallest footprint |
The cache costs 256 KiB per token at 16-bit, so the full 262,144-token context of one request needs about 64.0 GiB. Concurrency multiplies that number, not the weights.
That figure is an upper bound: this model uses recurrent layers mixed with attention layers, so only some layers keep a cache.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | 4 | 25.4 GiB | about 104,219 tokens |
| RTX 5090 32GB | 2 | 1.6 GiB | about 6,734 tokens |
| L40S 48GB | 2 | 30.4 GiB | about 124,699 tokens |
| A100 80GB | 1 | 18.5 GiB | about 75,956 tokens |
| H100 80GB | 1 | 18.5 GiB | about 75,956 tokens |
| RTX PRO 6000 96GB | 1 | 32.9 GiB | about 134,939 tokens |
| H200 141GB | 1 | 73.4 GiB | the full 262,144 tokens |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve sabicap/sabi-noema-mla \ --max-model-len 102400 \ --gpu-memory-utilization 0.90
Questions
The weights take about 51 GiB at the released precision. That fits on one A100 80GB, which leaves roughly 19 GiB for the key-value cache.
The repository declares apache-2.0. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 262,144 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email [email protected] with the model name and roughly what volume you expect.
Also in the library
Hy-Memory-30B-A3B is a Tencent-developed long-term-memory writer model designed to replace LLMs in Mem0, MemOS and Hy-Memory fram…
Granite Switch 4.1 30B is a modular LLM combining a 30-billion base, decision and guardian adapters, and an English speech-to-tex…
LFM2.5-24B-A2B is a 24 billion-parameter MoE reasoning model with 2.45 billion active parameters for on-device agentic tasks.
LFM2.5-24B-A2B-Base is a 24 billion-parameter mixture-of-experts model from LiquidAI aimed at on-device deployment and heavy fine…
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.