Skip to content
DocsGet Noema

Model ·

Noema 2B, our first model

A 2-billion-parameter model post-trained from Qwen3.5-2B for accurate, efficient reasoning on personal hardware.

By NoemaAI2 min read

Noema 2B has since been superseded by Noema 1.5 2B, released on August 12, 2026, which builds on it.

Noema 2B is NoemaAI's first model: a compact, instruction-tuned language model built on Qwen3.5-2B and post-trained for efficient local inference. It was trained and evaluated in non-thinking mode and is designed for a practical balance of quality and memory use. It gains most on the skills it was trained for (math, code and instruction following), is roughly flat on general reasoning, and gives up some factual recall.

Results against Qwen3.5-2B

Both models were scored in non-thinking mode with identical prompts. These comparisons were not reported with significance tests, and subsets of 100 to 200 items carry a standard error of about 4 to 5 points.

Frozen Noema harness, non-thinking mode (transformers 5.12.1, BF16, seed 1234). MATH-500 at a 2,048-token budget; IFEval on a 100-prompt subset. Percentages; difference in points.
BenchmarkNoema 2BQwen3.5-2BDifference
MATH-500, 2,048-token budget66.255.4+10.8
GSM8K, single sample82.579.2+3.3
GSM8K, self-consistency over 891.784.2+7.5
IFEval, prompt strict7365+8
Standard EvalPlus harness, non-thinking, pass@1, with Q8_0 GGUF builds of both models. Not directly comparable with the table above.
BenchmarkNoema 2BQwen3.5-2BDifference
HumanEval53.739.0+14.7
HumanEval+50.036.0+14.0
MBPP63.560.6+2.9
MBPP+50.048.4+1.6

Shorter, steadier answers

  • On IFEval, Noema 2B uses 394 tokens per correct answer against 565 for its base, about 30% fewer, and wrong answers also stop sooner instead of running to the limit.
  • At half the MATH-500 token budget it keeps 95.7% of its full-budget accuracy, against 82.4% for its base.
  • In our runaway-generation test, 2 of 120 answers never terminated. Other wrong answers end cleanly, with a parseable answer.

Where it falls short

  • A small knowledge cost against the base: MMLU-Pro −1.8, and about −4.5 on the older 57-subject MMLU.
  • Non-thinking mode only. Thinking mode is not supported and is prone to reasoning that never ends at this size.
  • Several results use 100- to 120-item subsets, with a standard error of about 4 to 5 points.
  • One architecture family and one base model; generality beyond this setting is untested.

Specifications

  • About 2B parameters, post-trained from Qwen3.5-2B.
  • Hybrid architecture: Gated DeltaNet and attention, 24 layers, hidden size 2,048, with a 262K-token context.
  • Non-thinking mode, trained and evaluated in English, text only.
  • SafeTensors (BF16) and GGUF (F16, Q8_0, Q6_K and Q4_K_M).

How to run it

Noema 2B came to the Noema app with Noema 3.1 on July 8, 2026. To run the GGUF build with llama.cpp, the model card suggests Qwen's non-thinking settings:

llama.cpp
llama-cli -m Noema-2B-Q4_K_M.gguf --jinja -cnv -st \
  -p "A farmer has 17 red apples and 25 green apples. How many in total?" \
  -n 256 --temp 0.7 --top-p 0.8 --top-k 20 --presence-penalty 1.5

License

Noema 2B is an open-weight release, not open source: the weights are publicly downloadable, but no open-source license is granted. Full results are on the model page.