---

base_model:

- NoemaAI-labs/Noema-1.5-2B

language:

- en

pipeline_tag: text-generation

tags:

- noema

- qwen3.5

- gguf

- llama.cpp

- local-ai

- reasoning

- code

- instruction-following

- multi-turn

---

Noema 1.5 2B GGUF

GGUF releases of Noema 1.5 2B, an open-weight local-first language model with recovered knowledge, stronger code generation, substantially better strict instruction following, improved multi-turn reliability, and more efficient reasoning.

This is an open-weight release, not an open-source release. The weights are publicly downloadable, but no open-source license is granted with this repository at this time.

These files were converted from the frozen merged BF16 release candidate with upstream llama.cpp revision 2d973636e292ee6f75fadcf08d29cb33511f509f. MTP is disabled. All files embed the public name, the original chat template, EOS token 248046 (<|im_end|>), PAD token 248044 (<|endoftext|>), and the 262,144-token architecture limit.

Files

| File | Quantization | Size | SHA-256 | Suggested use |

|---|---|---:|---|---|

| Noema-1.5-2B-BF16.gguf | BF16, unquantized | 3.52 GiB | 29e45221029eb396aeb493b549e91f5c29337c25fabe75259a67bf1262b2cb09 | Archival and maximum fidelity |

| Noema-1.5-2B-Q8_0.gguf | Q8_0 | 1.87 GiB | c7ef0f28f22b2ffd970c8f36b937f17e9410f98b2d9823207be7da7a162e2eda | Near-lossless deployment |

| Noema-1.5-2B-Q6_K.gguf | Q6_K | 1.45 GiB | 4895f1695f0e9cdc5bd138ff984942fc2b4f3accacf92ae74a49a1e47097607c | High-quality local use |

| Noema-1.5-2B-Q5_K_M.gguf | Q5_K_M | 1.31 GiB | 34a37539d76a1855b474767763bb7ea63a86713b1793564dd9bb8271ae863ac6 | Balanced quality and memory |

| Noema-1.5-2B-Q4_K_M.gguf | Q4_K_M | 1.19 GiB | e8e56badd716c533138342514d4175f861863107c938f73ca9db7c0469eccb3f | Recommended compact default |

Q5_K_M and Q4_K_M use the upstream mixed K-quant recipes, retaining selected sensitive tensors at higher precision.

Usage

With a current llama.cpp build:

llama-cli \
  -hf NoemaAI-labs/Noema-1.5-2B-GGUF:Q4_K_M \
  -cnv \
  -rea off \
  --temp 0

Non-thinking mode is recommended for concise instructions, structured generation, and code. The evaluated non-thinking preset is greedy: thinking off, temperature 0, top-p 1, and no top-k truncation.

For harder reasoning tasks, the evaluated thinking preset used thinking on, temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 1.5, and repetition penalty 1.0. Thinking can still become long or repetitive, so set an output limit.

The GGUF metadata advertises the Qwen3.5 backbone's native 262,144-token limit. Noema independently validated contexts only up to 24,576 tokens. Select context length according to available memory and workload.

Model summary

• Architecture: Qwen3.5 hybrid Gated DeltaNet and gated attention
• Parameters: approximately 2B
• Layers: 24
• Hidden dimension: 2,048
• Primary evaluated language: English
• Release status: open-weight; not open source
• Multimodality: text trunk only
• MTP: disabled

Benchmark details, training methodology, intended uses, and limitations are documented in the native model card.