Get Noema
← All models

Model 03 · Open weights · October 13, 2026

Noema 1.5 4B

Two reasoning styles of one 4B model. Deep thinks longer and is the more accurate; Fast answers in a fraction of the tokens and almost never gets stuck.

Get Noema Deep on Hugging Face Fast on Hugging Face
Parameters
4B
Base model
Qwen3.5-4B
Styles
Deep · Fast
Release
Open weights

Which one should I use

Same model size. Two ways to think.

Both are post-trained from Qwen3.5-4B, have the same size and run on the same devices. What separates them is how long they think.

Accuracy first

Noema 1.5 4B Deep

Thinks longer on hard problems and reaches its base model's accuracy with far fewer tokens.

Best for
Hard mathematics and competitive programming, when an answer is worth waiting for. Suits a Mac.
Strongest at
Thinking at temperature 0.6, where it is at or above Qwen3.5-4B on all seven benchmarks, and 4-bit robustness.
Trade-off
Under Qwen's default preset it trails Qwen3.5-4B on AIME and HMMT. About 14% of hard stress-test prompts run past 32K tokens.
Typical answer
14K tokens on GPQA Diamond, 23K on AIME 2026.
Output limit
81,920 tokens for contest math and code, 32,768 otherwise.
Hugging Face

Speed first

Noema 1.5 4B Fast

Answers in a fraction of the tokens and almost never gets stuck, even on hard prompts or as a 4-bit build.

Best for
Everyday questions, science and writing, on phones or with tight output budgets.
Strongest at
Loop resistance (1.1% against 65.0% on the hard stress test), GPQA Diamond at a third of the tokens, and instruction following.
Trade-off
Competition math: about 16 points behind Qwen3.5-4B on AIME and HMMT under Qwen's preset, and 8.3 behind on AIME at temperature 0.6.
Typical answer
9K tokens on GPQA Diamond, 16K on AIME 2026.
Output limit
32,768 tokens.
Hugging Face

Fast for everyday questions and phones. Deep for hard problems when you can wait.

Benchmarks

Noema 1.5 4B Deep

At temperature 0.6, matches or beats Qwen3.5-4B on every thinking benchmark we ran, with 1.2 to 3.1 times fewer tokens.

  • 1.2–3.1×

    Fewer tokens, same or better score

    At temperature 0.6, at or above Qwen3.5-4B on all seven thinking benchmarks. The largest gains are significant; AIME and HMMT are within noise.

  • 80.0%

    AIME 2026 at temperature 0.6

    Against 77.3 for Qwen3.5-4B (+2.7, within noise), at 22.7K tokens per answer instead of 36.9K.

  • 74.7%

    GPQA Diamond within 32K tokens

    Under Qwen's default preset: level with the 75.9 Qwen3.5-4B reaches only with 81,920 tokens, at about half the tokens.

  • 0.0% loops at 81,920

    Long reasoning that finishes

    All 139 stress-test answers Deep was still writing at 32,768 tokens finished within 81,920 (greedy decoding). Qwen3.5-4B kept repeating on about 58% of the test.

  • Noema 1.5 4B Deep
  • Noema 1.5 4B Fast
  • Qwen3.5-4B
  • Qwen-reported*

AIME 2026

Competition math

Thinking · temperature 0.6 · avg@16 · 81,920 tokens

HMMT Feb 2026

Competition math

Thinking · temperature 0.6 · avg@16 · 81,920 tokens

LiveCodeBench v6

Competitive programming

Thinking · temperature 0.6 · avg@4 · 81,920 tokens

GPQA Diamond

Graduate-level science

Thinking · temperature 0.6 · avg@8 · 32,768 tokens

  • Same harness, prompts and graders for every model; BF16 weights served by vLLM; thinking mode at temperature 0.6, top-p 0.95, top-k 20. Figures under the bars are mean output tokens per answer.
  • Unfinished: answers still being written at the output limit, scored as wrong for every model. Qwen3.5-4B's share is printed under its bar where it is 10% or more.
  • *Hatched: Qwen's published scores for Qwen3.5-4B (55.8 on LiveCodeBench v6, 76.2 on GPQA Diamond). We reproduce them under Qwen's own settings: 57.1 on LiveCodeBench with its default preset, and 75.9 on GPQA with 81,920 tokens.
  • Under Qwen's default preset (temperature 1.0, presence penalty 1.5), Qwen3.5-4B scores higher on AIME (85.4 vs 80.4) and HMMT (64.0 vs 60.2); both differences are significant.

Efficiency

Accuracy per token

Mean of AIME 2026, HMMT Feb 2026, LiveCodeBench v6 and GPQA Diamond, against mean output tokens per answer.

  • Noema 1.5 4B Deep
  • Noema 1.5 4B Fast
  • Qwen3.5-4B

Qwen's default preset

Mean output tokens per answer (fewer is better)

Noema 1.5 4B Deep
67.2% · 22.1K tokens
Noema 1.5 4B Fast
61.0% · 14.4K tokens
Qwen3.5-4B
66.4% · 37.4K tokens · 20% unfinished

Temperature 0.6

Mean output tokens per answer (fewer is better)

Noema 1.5 4B Deep
66.8% · 22.8K tokens
Noema 1.5 4B Fast
62.3% · 15.1K tokens
Qwen3.5-4B
55.7% · 41.8K tokens · 37% unfinished

Under Qwen's default preset, Deep reaches its base model's mean accuracy with about 40% fewer tokens, while Fast gives up about five points of that mean for answers less than half as long. At temperature 0.6, where Qwen3.5-4B often runs out of room, both are well ahead of it at a fraction of its length.

  • Same harness and prompts for every model; BF16 weights; thinking mode; benchmarks weighted equally. Unfinished: answers still running at the output limit, scored as wrong.
  • Qwen's default preset: temperature 1.0, top-p 0.95, top-k 20, presence penalty 1.5. Output limit 81,920 tokens for AIME, HMMT and LiveCodeBench, 32,768 for GPQA.
  • Qwen reports 55.8 (LiveCodeBench) and 76.2 (GPQA) for Qwen3.5-4B. We reproduce both under Qwen's settings: 57.1 on LiveCodeBench with its default preset, 75.9 on GPQA with 81,920 tokens.
  • Per-benchmark results differ: under Qwen's default preset, Qwen3.5-4B scores higher on AIME and HMMT.

Reliability

A 4B model that knows when to stop.

Small reasoning models often get stuck: they repeat a passage, or keep reasoning until they run out of room, and never answer.

On 1,000 hard held-out prompts, Qwen3.5-4B does this on 65% of them under greedy decoding. Noema 1.5 4B Fast does it on 1.1%, a reduction of 98% or more under every setting we tested; Deep's reduction is 70 to 79%.

Deep reasons longer, and on 14% of those prompts it is still working at 32,768 tokens. Given the 81,920 tokens Qwen recommends for hard problems, every one of those answers finishes, while Qwen3.5-4B is still repeating itself on an estimated 58%. In 100 long conversations of up to 50 turns, neither Noema model looped once; Qwen3.5-4B looped in 51.

  • Noema 1.5 4B Deep
  • Noema 1.5 4B Fast
  • Qwen3.5-4B

Greedy, 32K limit

Doom-loop rate

Thinking · greedy · 32,768 tokens · lower is better

4-bit on llama.cpp

Doom-loop rate

Q4_K_M · llama.cpp defaults · 32,768 tokens · lower is better

Greedy, 81,920 limit

Still looping with the full budget

Thinking · greedy · failures rerun · lower is better

Long conversations

Conversations with a loop, of 100

Thinking · temperature 1.0 · 20 to 50 turns · lower is better

  • Doom loop: a passage of 60 or more characters repeated four or more times, or an answer still running at 32,768 tokens. The stress test is 1,000 held-out prompts (400 hard math, 200 science, 200 code, 100 open-ended, 100 multi-turn), so its rates are not everyday rates.
  • Every reduction is significant: for example, greedy loops fall by 63.9 points for Fast (95% CI 60.8 to 66.8) and 51.1 for Deep (48.0 to 54.2).
  • The 81,920-token panel reran every greedy failure with the larger limit, counting prompts already finished as finished; only greedy decoding was rerun. For Qwen3.5-4B it reran a fixed random 200 of its 650 failures (95% CI 54.9 to 60.5%).
  • Long conversations: 100 conversations of 20 to 50 turns, every turn generated by the model under test. 4-bit rows use each model's own Q4_K_M build under llama.cpp default sampling.

Full results

Every benchmark, with the tokens it took.

Score and mean output tokens per answer, with each Noema model's paired difference from Qwen3.5-4B on the same items.

Our recommended setting and Qwen's suggestion for precise coding: temperature 0.6, top-p 0.95, top-k 20, no presence penalty. Here Qwen3.5-4B often runs out of room; its unfinished answers count as wrong.

Thinking · temperature 0.6. Our recommended setting and Qwen's suggestion for precise coding: temperature 0.6, top-p 0.95, top-k 20, no presence penalty. Here Qwen3.5-4B often runs out of room; its unfinished answers count as wrong.
BenchmarkNoema 1.5 4B DeepDeepNoema 1.5 4B FastFastQwen3.5-4BQwen3.5-4B
AIME 202630 × 16 · 81,920 tokens80.022.7K+2.7−1.7 to +6.7, 95% CI −1.7 to +6.7, within noise69.015.6K−8.3−16.0 to −1.2, 95% CI −16.0 to −1.2, significant loss77.336.9K15% unfinished
HMMT Feb 202633 × 16 · 81,920 tokens55.331.8K+2.1−3.2 to +8.1, 95% CI −3.2 to +8.1, within noise51.718.4K−1.5−7.6 to +5.3, 95% CI −7.6 to +5.3, within noise53.250.3K34% unfinished
arxivmath 082657 × 16 · 81,920 tokens3.339.3K+3.0+1.2 to +5.0, 95% CI +1.2 to +5.0, significant gain4.120.1K+3.7+1.2 to +6.7, 95% CI +1.2 to +6.7, significant gain0.378.1K89% unfinished
LiveCodeBench v6175 × 4 · 81,920 tokens55.923.4K+19.0+14.6 to +23.9, 95% CI +14.6 to +23.9, significant gain52.917.3K+16.0+11.6 to +20.7, 95% CI +11.6 to +20.7, significant gain36.956.4K59% unfinished
GPQA Diamond198 × 8 · 32,768 tokens75.813.6K+20.4+16.4 to +24.6, 95% CI +16.4 to +24.6, significant gain75.69.2K+20.2+16.0 to +24.6, 95% CI +16.0 to +24.6, significant gain55.423.4K38% unfinished
MMLU-Pro12,032 × 1 · 32,768 tokens79.64.8K+0.5+0.01 to +1.1, 95% CI +0.01 to +1.1, significant gain79.23.8K+0.2−0.3 to +0.7, 95% CI −0.3 to +0.7, within noise79.05.9K
IFEval541 × 1 · prompt strict · 32,768 tokens94.14.8K+36.2+32.0 to +40.5, 95% CI +32.0 to +40.5, significant gain92.24.1K+34.4+30.1 to +38.6, 95% CI +30.1 to +38.6, significant gain57.915.0K38% unfinished
Mean of the sevendifference from Qwen3.5-4B+12.0+10.5 to +13.5, 95% CI +10.5 to +13.5, significant gain+9.2+7.4 to +11.1, 95% CI +7.4 to +11.1, significant gainbaseline

Green: significant gain over Qwen3.5-4B. Red: significant loss. Grey: within noise, because the 95% interval includes zero. Bold marks the best score in each row.

What each benchmark measures
AIME 2026
Competition mathematics from the American Invitational Mathematics Examination30 problems · Share answered correctly
HMMT Feb 2026
Competition mathematics from the Harvard-MIT Mathematics Tournament33 problems · Share answered correctly
arxivmath 0826
Research-level mathematics questions drawn from recent arXiv papers57 problems · Share answered correctly
LiveCodeBench v6
Competitive programming; code is run against hidden tests175 problems · Share whose code passes every test
GPQA Diamond
Graduate-level multiple-choice science in physics, chemistry and biology198 questions · Share answered correctly
MMLU-Pro
Broad academic and professional knowledge, ten-option multiple choice12,032 questions · Share answered correctly
IFEval
Following verifiable formatting instructions such as word counts, sections and case541 prompts · Prompt-level strict: every instruction must be followed
IFBench
Following new, unseen kinds of verifiable instruction300 prompts · Prompt-level strict accuracy
Multi-IF
Instruction following across three-turn conversationsThree-turn conversations · Prompt-level loose, averaged over the turns
Loop stress test
Whether the model gets stuck: 400 hard math, 200 science, 200 code, 100 open-ended and 100 multi-turn prompts1,000 prompts · Share of answers that loop or are unfinished at 32,768 tokens
  • All three models ran in the same harness with identical prompts and graders, as BF16 weights served by vLLM 0.30.0, except the 4-bit rows (each model's own Q4_K_M build under llama.cpp default sampling). None of these sets was used for training or for choosing checkpoints.
  • Rows show items × samples per item and the output limit, which follows Qwen's guidance: 81,920 tokens for competition math and programming, 32,768 for other thinking tasks. An answer still running at the limit is unfinished and scored as wrong; token means count it at the limit, so Qwen3.5-4B's means understate what it would have spent.
  • Differences are paired: each Noema model's score minus Qwen3.5-4B's on the same items, with a 95% confidence interval from 10,000 bootstrap resamples. We call a difference a gain or a loss only when its interval excludes zero. AIME (30 problems) and HMMT (33) are small, so their intervals span several points.

Our runs and Qwen's published figures

Where our Qwen3.5-4B scores differ, its answers ran out of room.

Qwen and OpenBMB publish scores for Qwen3.5-4B from their own harnesses. Under comparable settings our measurements match them; where ours sit lower, Qwen3.5-4B left many answers unfinished at the output limit.

Qwen3.5-4B: our measurements against published figures
BenchmarkQwenOpenBMBOurs, Qwen's presetOurs, temperature 0.6Why they differ
MMLU-Pro79.178.079.179.0Matches.
IFEval89.890.290.857.9Matches under Qwen's preset. At temperature 0.6, 38% of answers are unfinished.
LiveCodeBench v655.856.457.136.9Matches under Qwen's preset. At temperature 0.6, 59% of answers are unfinished at 81,920 tokens.
GPQA Diamond76.277.1*59.155.4At 32,768 tokens, 33% and 38% of answers are unfinished. With 81,920 tokens we measure 75.9.
AIME 2026–82.785.477.3Close to OpenBMB's figure.
HMMT Feb 2026–64.064.053.2Matches OpenBMB's figure.
  • *OpenBMB's GPQA Diamond figure is quoted from Artificial Analysis. Given 81,920 tokens under Qwen's default preset, Qwen3.5-4B scores 75.9 at 28,915 tokens per answer; paired against it, Fast's 32,768-token score differs by −0.6 points (95% CI −3.0 to +2.0) and Deep's by −1.3 (−3.4 to +1.0), level within noise at 9,385 and 14,085 tokens per answer.

Run it locally

Open weights for hardware you own.

Download either style into the Noema model library and keep every prompt and response on hardware you control. Fast suits iPhone; Deep suits Mac. Each is a separate download. Weights are open; no open-source license is granted.

Status
Available · October 13, 2026
Parameters
4.21B
Base model
Qwen3.5-4B
Styles
Deep · Fast
Architecture
Hybrid: Gated DeltaNet + gated attention
Layers / hidden
32 / 2,560
Context
262,144 tokens (architecture limit); outputs evaluated to 81,920
Modes
Thinking (default) · non-thinking
Sampling
Temperature 0.6 · top-p 0.95 · top-k 20
Output limit
32,768; Deep 81,920 for contest math and code
Training
On-policy distillation, GRPO with a finishing reward, FTPO
Formats
SafeTensors · GGUF (BF16 to Q4_K_M) · MLX (BF16 to 4-bit)
Download
About 2.6 GB each at Q4_K_M
Language
English, text only (evaluated)
Release
Open weights; not open source

Evaluation discipline

Paired runs, Qwen's own limits, losses included.

Every comparison against Qwen3.5-4B is paired on identical items, with the same prompts, answer extraction, graders, sampling settings and output limits. The benchmark panel and the loop stress test were held out from training and checkpoint selection.

Paired, with intervals
95% CIEach difference comes with a 95% interval from 10,000 bootstrap resamples. We call it a gain or a loss only when the interval excludes zero.
GPQA Diamond budget
75.9 at 81,920At 32,768 tokens a third of Qwen3.5-4B's answers are still running. Given 81,920 tokens it scores 75.9, matching Qwen's reported 76.2.
Long reasoning, not looping
139 of 139Every hard prompt Deep was still working on at 32K tokens finished once allowed 81,920 (greedy decoding).

Evaluation controls

  • One harness for every model: same prompts, answer extraction and graders
  • BF16 weights served by vLLM 0.30.0; 4-bit rows use Q4_K_M under llama.cpp
  • Qwen's recommended output limits; an answer still running at the limit counts as wrong
  • AIME, HMMT and arxivmath avg@16; LiveCodeBench avg@4; GPQA avg@8
  • Paired 95% intervals from 10,000 bootstrap resamples of the items
  • Training data 13-gram decontaminated against every evaluation set

Measured limits

  • Under Qwen's default preset, both models trail Qwen3.5-4B on AIME 2026 and HMMT Feb 2026, and Fast also on LiveCodeBench v6.
  • Fast trails on competition math in every setting: about 16 points on AIME and HMMT under Qwen's preset, 8.3 on AIME at temperature 0.6.
  • Without thinking the models are close; Fast is lower on LiveCodeBench v6 (−4.9, significant).
  • Loop rates come from a stress test weighted toward hard prompts; they are not everyday rates.
  • 4-bit results cover the loop stress test and non-thinking GPQA and IFEval only.
  • One training run per tier; evaluated in English, text only, with outputs up to 81,920 tokens.