Math at a deployment budget
66.2 vs 55.4 for the base at a 2,048-token budget, the setting small models actually run in.
Noema model family
Noema 2B
A 2-billion-parameter model designed for accurate, efficient local inference on personal hardware.
Why it exists
Built to reason within a real device budget.
Noema 2B is our first release: a compact, instruction-tuned language model built on Qwen3.5-2B and post-trained for more efficient local inference. It runs on your own hardware in non-thinking mode and is designed for a practical balance of quality and memory use. On held-out verifiable prompts, it performs comparably to an open model more than four times its size.
- 01Token-efficient reasoning
- 02Math & step-by-step reasoning
- 03Instruction following
- 04On-device, non-thinking
- Status
- Available now
- Released
- July 6, 2026
- Parameters
- 2.0B
- Base model
- Qwen3.5-2B
- Architecture
- Hybrid: Gated DeltaNet + attention
- Context length
- 262K tokens
- Modes
- Non-thinking · thinking
- Formats
- SafeTensors · GGUF
- Precision
- F16 · Q8_0 · Q6_K · Q4_K_M
- License
- Apache 2.0
Load Noema 2B locally, choose a device-aware runtime preset, and keep inference on hardware you control.
Post-training results
What changed after the base model.
Measured improvements on the frozen Noema evaluation harness.
MATH-500 @2048
Grade-school math
+3.3 single-shot and +7.5 with self-consistency; pass@8 reaches 96.7%.
Instruction following
Prompt-strict 73 vs 65, with gains across all four IFEval metrics and 22% fewer tokens per response.
Code that runs
HumanEval 53.7 vs 39.0 and HumanEval+ 50.0 vs 36.0. The result is markedly more correct, runnable code than the base.
Fewer tokens per answer
394 vs 565 tokens per correct answer on IFEval; wrong answers also fail shorter instead of running to the cap.
Reliable generation
Runaway, repetitive generation is effectively gone. Failures terminate with clean, parseable answers.
Frozen evaluation harness
Benchmarks
Higher is better · non-thinking mode
84%
Performance at a smaller size
On 130 held-out verifiable prompts never seen in training, Noema 2B matches or beats Qwen3.5-9B on 84% of items. That open model is over four times its size.
95.7%vs 82.4% for the base
Budget-robust
Share of full-budget MATH-500 accuracy retained at half the token budget. Noema stays near its ceiling; the base loses nearly a fifth of its performance when tokens are tight.
Versus the base and a 9B reference
05 measures
GSM8K
GSM8K
MATH
MBPP-100
IFEval
All three scored on the same frozen harness in non-thinking mode, with an identical evaluation for each. Qwen3.5-9B, 4.5× Noema's size, is a reference ceiling: Noema 2B closes much of the base → 9B gap.
Coding versus the base
04 measures
HumanEval
HumanEval+
MBPP
MBPP+
Noema 2B vs Qwen3.5-2B on our frozen harness in non-thinking mode at pass@1. EvalPlus suite: HumanEval / MBPP plus their extended-test (+) variants. Current general models publish LiveCodeBench, not EvalPlus, so there is no fair current-peer set here.
Versus the current 2–4B class
03 measures
IFEval
GSM8K
MATH-500
Non-thinking mode for all. LFM2-2.6B comes from the LFM2 technical report (arXiv:2511.23404; no reasoning mode). SmolLM3-3B comes from community /no_think evals: GSM8K flexible-extract and IFEval prompt-level loose (strict not published), with MATH-500 pending. Qwen3.5-4B non-thinking is pending. Noema uses our frozen harness, with MATH-500 at a 2,048-token budget. Protocols differ across labs.
Apache 2.0 / Open weights
Run Noema 2B on hardware you own.
Start with the official weights, or download a GGUF quant made for local runtimes.
View Noema 2B on Hugging Face