Known-issues sheets for local language models Rev 0.1.0 · reviewed every 7 days

Model datasheet · Nemotron 3

Nemotron 3 Nano 30B-A3B

The long-context specialist of this catalog by a distance: 86.3% on RULER at one million tokens, where most models here are unrated and the best measured rival manages 66% at 128K. A hybrid Mamba-Transformer MoE — only 6 of its 52 layers hold a KV cache, which is why the window is affordable rather than nominal.

Vendor
NVIDIA
Architecture
MoE · 31.6B total / ~3.6B active
Context
262,144 tokens
License
NVIDIA Nemotron Open Model License
Released
2025-12-15
Vision
No

→ GGUF (llama.cpp / LM Studio / Ollama) → MLX (Apple Silicon)

§1 Characteristics by quantization band

Q2–Q3 — 2–3 bit

19GB weights · e.g. Q3_K_M + KV: 48MB at 8K · 192MB at 32K · 768MB at 128K

— Not yet rated · contributions welcome —

Q4–Q5 — 4–5 bit

24.7GB weights · e.g. Q4_K_M , mlx-4bit + KV: 48MB at 8K · 192MB at 32K · 768MB at 128K
Long context 8/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: one point down. This is the band to run — 24.7GB, and the tiny KV cache means a very long context stays affordable on top of it.

Math & reasoning 7/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: one point down for 4-bit MoE experts on multi-step arithmetic.

Coding 5/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: one point down from the full band.

Q6–Q8 — 6–8 bit

33.4GB weights · e.g. Q6_K , Q8_0 + KV: 48MB at 8K · 192MB at 32K · 768MB at 128K
Long context 9/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q6_K · aggregated · 2026-08-30

    Editorial derivation: long-context retrieval is not quant-sensitive in the way arithmetic is, and 6-bit is near-lossless here. Carried unchanged.

Math & reasoning 8/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q6_K · aggregated · 2026-08-30

    Editorial derivation: carried from the full band.

FP16 — full precision (fp16/bf16)

63.2GB weights · e.g. bf16 + KV: 48MB at 8K · 192MB at 32K · 768MB at 128K
Long context 9/10
data checked Aug 2026
  • vendor-model-cardRULER-100@1M = 86.3 @ bf16 · vendor-reported · 2026-08-30

    92.9 at 256K and 91.3 at 512K. The best long-context evidence on this site by a wide margin — Gemma 4 31B, the nearest measured rival, manages 66.4% at 128K. RULER is a multi-hop, multi-needle test, not a single planted sentence, which is what makes the number meaningful.

Math & reasoning 8/10
data checked Aug 2026
  • vendor-model-cardAIME25 = 89.1 @ bf16 · vendor-reported · 2026-08-30

    GPQA 73.0, MMLU-Pro 78.3. Scored without tools; NVIDIA also reports 99.2 on AIME25 with tools, which measures the harness as much as the model.

Coding 6/10
data checked Aug 2026
  • vendor-model-cardLiveCodeBench = 68.3 @ bf16 · vendor-reported · 2026-08-30

    SWE-Bench (OpenHands): 38.8 — a long way below the competition-style number, and the honest signal for real repository work. This is not the coding model in this catalog.

Tool calling / agentic 5/10
data checked Aug 2026
  • vendor-model-cardBFCL v4 = 53.8 @ bf16 · vendor-reported · 2026-08-30

    Modest for a model marketed as agentic, and measured on BFCL v4 rather than the v3 most other entries here quote — not directly comparable.

Instruction following 7/10
data checked Aug 2026
  • vendor-model-cardIFBench = 71.5 @ bf16 · vendor-reported · 2026-08-30

    One of only eleven models on this site with any instruction-following evidence at all.

§2 Known issues & what fixes them

ISSUE-01 severe Q2–Q3 · math, coding, tool-calling

Quantizations below 4-bit are reported broken on this model, not merely degraded. The architecture is the reason: several tensor dimensions are not divisible by 128, so parts of the model cannot be quantized to lower bit widths at all — which is also why Q6_K (33.43GB) and Q8_0 (33.51GB) come out almost the same size.

Fix
Move up a quant band [strong]

Use Q4_K_M or better. The 3-bit build saves 5.7GB and costs far more than that in quality.

Evidence · 2 sources · community-consensus
ISSUE-02 moderate Q2–Q3 / Q4–Q5 / Q6–Q8 / FP16 · coding

LiveCodeBench 68.3 looks strong; SWE-Bench with OpenHands is 38.8. The gap between competition-style problems and real repository work is wide here, and the second number is the one that predicts whether it can fix your bug.

No fix
No known fix — choose a different model

Pick it for long documents and reasoning, not for agentic coding. Ornith-1.5 35B-A3B and Qwen3.6 27B are the coding models at this footprint.

Evidence · 1 source · anecdotal
ISSUE-03 mild Q2–Q3 / Q4–Q5 / Q6–Q8 / FP16 · instruction-following

Released under the NVIDIA Nemotron Open Model License rather than Apache-2.0 or MIT. The weights are downloadable and commercial use is permitted, but it is a vendor licence with its own terms — read them before shipping something on top of it.

No fix
No known fix — choose a different model

Not a defect. If a permissive licence is a hard requirement, Qwen3.6 35B-A3B (Apache-2.0) and Ornith-1.5 35B-A3B (MIT) are the comparable alternatives.

Evidence · 1 source · anecdotal