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

Model datasheet · Nanbeige4.2

Nanbeige4.2 3B

A looped transformer: 22 layers run twice, so 3B of parameters do the work of a much deeper model. Reported to beat Qwen3.5 9B and Gemma 4 12B on agentic coding at a third of the size. The loop is not free — it doubles the KV cache, which is the largest memory cost here by far.

Vendor
Nanbeige
Architecture
Dense · 3B
Context
262,144 tokens
License
Apache-2.0
Released
2026-07-21
Vision
No

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

§1 Characteristics by quantization band

Q2–Q3 — 2–3 bit

2.2GB weights · e.g. Q3_K_M + KV: 1.4GB at 8K · 5.5GB at 32K · 22GB at 128K
Coding 4/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q3_K_M · aggregated · 2026-08-30

    Editorial derivation: three points down. 3-bit on a 3B model saves 0.44GB over 4-bit and costs a great deal — see the known issue.

Math & reasoning 3/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q3_K_M · aggregated · 2026-08-30

    Editorial derivation: the standard collapse of multi-step arithmetic at ≤3-bit, worse than usual on a model this small.

Q4–Q5 — 4–5 bit

2.7GB weights · e.g. Q4_K_M , mlx-4bit + KV: 1.4GB at 8K · 5.5GB at 32K · 22GB at 128K
Coding 6/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: one point down. A 3B at 4-bit is a thin margin; the looped architecture is also unusual enough that community quant evidence for it is scarce.

Math & reasoning 6/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: two points down from the full band.

Q6–Q8 — 6–8 bit

3.6GB weights · e.g. Q6_K + KV: 1.4GB at 8K · 5.5GB at 32K · 22GB at 128K
Coding 7/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q6_K · aggregated · 2026-08-30

    Editorial derivation: carried from the full band. At 3.6GB the weights fit anything.

Math & reasoning 7/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q6_K · aggregated · 2026-08-30

    Editorial derivation: one point down — small models have the least redundancy to lose, and maths goes first.

FP16 — full precision (fp16/bf16)

8.3GB weights · e.g. bf16 + KV: 1.4GB at 8K · 5.5GB at 32K · 22GB at 128K
Coding 7/10
data checked Aug 2026
  • vendor-model-cardSWE-bench Verified = 63.6 @ bf16 · vendor-reported · 2026-08-30

    Against Qwen3.5 9B at 53.1 and Gemma 4 12B at 44.2 on the same table — a 3B beating models three and four times its size. Published with an arXiv technical report, which is more than most vendor claims on this site carry, but still the authors' own runs.

Math & reasoning 8/10
data checked Aug 2026
  • vendor-model-cardGPQA Diamond = 87.4 @ bf16 · vendor-reported · 2026-08-30

    HMMT Feb 2026: 82.8. Extraordinary for 3B — higher than several 27B-class models here. Treat as a claim pending independent replication; a figure this far out of line with size is exactly what benchmark contamination looks like.

§2 Known issues & what fixes them

ISSUE-01 severe Q2–Q3 / Q4–Q5 / Q6–Q8 / FP16 · long-context

The looped architecture runs its 22 layers twice, and each pass caches its own keys and values — so the KV cache behaves like a 44-layer model even though only 3B of weights are stored. At 32K context that is roughly 5.5GB of cache against 2.7GB of weights: the context costs twice the model. The usual intuition that a 3B is cheap to run holds for the download and fails badly for the session.

Fix
RAG (retrieval) [strong]

Keep the context short and retrieve what you need. On this model that is a memory decision, not just a quality one.

Fix
Update your runtime [partial]

Loop-aware KV handling is new; check whether your runtime caches once per layer or once per pass, because the memory figures here assume the latter (the safe assumption).

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

GPQA-Diamond 87.4 and SWE-bench Verified 63.6 from a 3B model would place it above most 27B-class models in this catalog. That is either a genuine architectural result or contamination, and from the vendor's own runs there is no way to tell. The arXiv report is a point in its favour; independent replication is what would settle it.

No fix
No known fix — choose a different model

No fix — a caveat. Try it on your own work before believing the table; a 2.7GB download makes that cheap.

Evidence · 2 sources · community-consensus