Open weights · Released August 12, 2026
Noema 1.5 2B
A compact open-weight 2B model that follows instructions more reliably and reasons in fewer tokens than stock Qwen3.5-2B, while keeping its broad knowledge.
- Parameters
- About 2B
- Base model
- Noema 2B
- Formats
- SafeTensors · GGUF · MLX
Why 1.5 exists
More reliable instructions, shorter reasoning.
In a one-time final evaluation against stock Qwen3.5-2B, Noema 1.5 2B follows strict instructions 7.0 points more often and uses 13.5% fewer thinking tokens, while passing a preregistered knowledge-retention test.
- +7.02 points
IFEval, prompt strict
72.09 against 65.06 for stock Qwen3.5-2B on 541 prompts; statistically significant (p = 0.000475).
- +3.11 points
Multi-IF, all three turns
17.88% of 4,501 conversations satisfied every turn, against 14.77%; statistically significant (p = 4.44e−8).
- +3.05 points
HumanEval+, pass@1
53.66 against 50.61 on 164 problems; favorable, but not statistically resolved.
- −13.5% tokens
Shorter thinking
Mean completion length fell from 7,969 to 6,895 tokens on 300 prompts, with accuracy at 68.67 against 67.33.
Noema 2B developed useful strengths in mathematics, code and structured responses, but paid a measurable price in broad knowledge. Noema 1.5 2B was trained from it to recover that deficit without giving up those capabilities. In the final evaluation it follows instructions more precisely, keeps multi-turn constraints more often and reasons in fewer tokens than stock Qwen3.5-2B, and its MMLU-Pro score passed a preregistered non-inferiority test. No retrieval, external tools or inference-time answer repair were used in the reported results; the improvements live in the weights.
- Precise instruction following
- Multi-turn constraint retention
- Shorter reasoning
- Broad knowledge retained
Benchmarks
Final evaluation against stock Qwen3.5-2B
Percentages unless noted. The frozen release and stock Qwen3.5-2B used identical prompts, generation settings and graders.
One-time final evaluation
Nine measures, paired on the same prompts
Higher is better · lower is better for tokens
- Noema 1.5 2B
- Stock Qwen3.5-2B
- The frozen Noema 1.5 2B release and stock Qwen3.5-2B used identical prompts, generation settings and graders. MMLU-Pro was the preregistered knowledge-retention endpoint; point estimates are reported even when they were not statistically resolved.
▲ statistically significant gain · ≈ not statistically resolved, or within the preregistered margin · deltas without a mark were not reported with a significance test.
Development progress against Noema 2B
Paired development and selection results against the exact Noema 2B checkpoint training began from. They informed checkpoint selection, so they are not a third arm of the final evaluation and are not used in the highlights above.
Development results
Versus Noema 2B
Higher is better
- Noema 1.5 2B
- Noema 2B
- Development and selection sets, paired against the starting checkpoint. No significance tests were reported for these results.
Publisher-reported context
How Noema 1.5 2B compares with published results from other compact open models.
Each publisher ran its own evaluation, so these numbers are directional. Only stock Qwen3.5-2B was measured in Noema's paired protocol above.
Directional context, not a leaderboard. External scores are reported by each model publisher and were not produced in Noema's controlled harness. Prompt templates, reasoning modes, sampling, token budgets, benchmark revisions, quantization and graders may differ.
Smaller open models
Open models of about 1.2B parameters
Published results for LFM2.5-1.2B Thinking and EXAONE 4.0 1.2B in both of its modes. The protocol beside each model is part of the comparison.
| Model | Protocol | Source |
|---|---|---|
| Noema 1.5 2B | Noema final evaluation · prompt-strict IFEval | Noema final evaluation |
| LFM2.5-1.2B Thinking | Thinking · averaged IFEval | Liquid AI model card |
| EXAONE 4.0 1.2B | Non-reasoning | LG AI model card |
| EXAONE 4.0 1.2B | Reasoning | LG AI model card |
Noema reports higher MMLU-Pro and GPQA-Diamond results than LFM2.5-1.2B Thinking and narrowly exceeds EXAONE's non-reasoning results. EXAONE leads both knowledge measures in reasoning mode, and LFM leads on instruction following.
Larger edge models
Selected larger local and edge models
Publisher results from larger local and edge-oriented models, on the same score scale.
| Model | Protocol | Source |
|---|---|---|
| Noema 1.5 2B | Noema final evaluation · prompt-strict IFEval | Noema final evaluation |
| Phi-4 Mini | 0-shot CoT | Microsoft model card |
| Gemma 4 E2B | Publisher evaluation | Google model card |
| Ministral 3 3B | Reasoning | Mistral model card |
| Nemotron 3 Nano 4B | FP8 · reasoning on | NVIDIA model card |
| Qwen3.5-4B | Publisher evaluation | Qwen model card |
Noema's reported MMLU-Pro result slightly exceeds Phi-4 Mini's. Gemma, Ministral, Nemotron and Qwen retain advantages on their published measures.
Model card
Run and cite Noema 1.5 2B
How to run it, what it is for, where it falls short, and the terms it is released under.
Quickstart
1. Install the Noema app on iPhone, iPad or Mac: noemaai.com/download
2. Open Explore → Models and search for "Noema 1.5 2B".
3. Choose a GGUF or MLX build that fits your device.
4. Review the fit guidance for your device, then tap Download.
5. When the download finishes, open Stored and load the model.
Every prompt and response stays on hardware you control.Q4_K_M is the recommended compact default. Other precisions are listed under Downloads. For other runtimes, see Run Noema models.
Downloads
| Format | Precision | Size | Download |
|---|---|---|---|
| Transformers | BF16 | 3.76 GB | Hugging Face: Noema 1.5 2B, Transformers BF16 |
| GGUF | BF16 | 3.78 GB | Hugging Face: Noema 1.5 2B, GGUF BF16 |
| GGUF | Q8_0 | 2.01 GB | Hugging Face: Noema 1.5 2B, GGUF Q8_0 |
| GGUF | Q6_K | 1.56 GB | Hugging Face: Noema 1.5 2B, GGUF Q6_K |
| GGUF | Q5_K_M | 1.41 GB | Hugging Face: Noema 1.5 2B, GGUF Q5_K_M |
| GGUF | Q4_K_M | 1.27 GB | Hugging Face: Noema 1.5 2B, GGUF Q4_K_MRecommended |
| MLX | BF16 | 3.76 GB | Hugging Face: Noema 1.5 2B, MLX BF16 |
| MLX | 8-bit | 2.00 GB | Hugging Face: Noema 1.5 2B, MLX 8-bit |
| MLX | 6-bit | 1.53 GB | Hugging Face: Noema 1.5 2B, MLX 6-bit |
| MLX | 4-bit | 1.06 GB | Hugging Face: Noema 1.5 2B, MLX 4-bit |
Sizes are the published weight files on Hugging Face, in gigabytes.
Specifications
- Parameters
- About 2B
- Base model
- Noema 2B
- Lineage
- Qwen3.5-2B → Noema 2B → Noema 1.5 2B
- Architecture
- Qwen3.5 hybrid: Gated DeltaNet linear attention with gated attention
- Layers / hidden size
- 24 / 2,048
- Context
- 262,144 tokens (architecture limit); independently validated to 24,576
- Modes
- Non-thinking (recommended) · thinking
- Training
- On-policy distillation from Noema 2B: Qwen3.5-9B scored most of the model's own samples, and untouched Qwen3.5-2B scored the rest to preserve general behavior. A rank-128 LoRA adapter, merged for release.
- Formats
- SafeTensors (BF16) · GGUF (BF16 to Q4_K_M) · MLX (BF16 to 4-bit)
- Language
- English (primary, evaluated), text only
- Released
Recommended settings
| Use | Thinking | Temperature | Top-p | Top-k | Presence penalty | Output limit |
|---|---|---|---|---|---|---|
| Instructions, structured output and code (recommended) | Off | 0 (greedy) | 1 | Off | 0 | Set a limit |
| Harder reasoning | On | 1.0 | 0.95 | 20 | 1.5 | Set a limit |
The packaged defaults are greedy and stop at <|im_end|>. Deployments should enforce output limits and detect repeated spans, especially with thinking on.
Intended use
- Local and privacy-sensitive assistants, including offline use.
- Structured and format-constrained generation.
- Multi-turn conversations that must keep earlier constraints.
- Short Python programs, education and lightweight reasoning.
- Deployments with tight memory and compute budgets.
Out of scope
- Knowledge-intensive work without retrieval or authoritative tools.
- Image or audio input, which was not trained or evaluated.
- Serving as the sole authority for high-stakes decisions.
Limitations
- GPQA-Diamond (−5.05) and IFBench (−3.67) finished below stock Qwen3.5-2B in the final point estimates.
- HumanEval+ and thinking accuracy were favorable but not statistically resolved.
- MMLU-Pro knowledge was retained within the preregistered margin, not improved.
- One primary training seed; the results do not show that this method beats matched supervised training.
- Thinking mode can still produce long or repetitive responses; enforce output limits.
- Evaluated primarily in English and text only; context validated to 24,576 tokens.
- Like any language model, it can state false information with confidence. Verify important facts, and do not rely on it alone for medical, legal, financial or safety-critical decisions.
Evaluation
The final comparison was run once, after the release candidate was frozen, with no score inspection before both arms were complete. No retrieval, external tools, answer repair or outside model calls were used to produce benchmark answers.
- Strict instructions
- p = 0.000475The +7.02-point IFEval gain was statistically significant.
- Complete multi-turn success
- p = 4.44e−8The +3.11-point all-three-turn gain was statistically significant.
- Knowledge retained
- 11,332 questionsMMLU-Pro passed the preregistered non-inferiority test: a one-sided 95% lower bound of −1.438 points against a −1.5-point margin.
- Frozen model and tokenizer revisions
- Identical prompts and settings across both arms
- BF16 inference with vLLM 0.19.1 on NVIDIA L40S hardware
- Pinned official graders
- Paired bootstrap intervals and McNemar tests
- Deterministic keys and exact-count verification
License
Open weights. The weights are publicly downloadable. This is an open-weight release, not open source: no open-source license is granted.
Citation
@misc{noema15_2b_2026,
title = {Noema 1.5 2B: Restoration-Aware On-Policy Distillation for Local AI},
author = {{NoemaAI}},
year = {2026},
month = aug,
howpublished = {\url{https://noemaai.com/models/noema-1-5-2b}},
note = {Noema 1.5 2B, open weights: https://huggingface.co/NoemaAI-labs/Noema-1.5-2B}
}