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

Model datasheet · Granite 4.2

Granite 4.2 8B

The tool-calling specialist of the 8GB tier: Apache-2.0, a thinking switch, and function-calling trained in live sandboxes rather than scraped transcripts. Its predecessor's 8B matched a 32B MoE across ten benchmarks — this is that lineage, five days old.

Vendor
IBM
Architecture
Dense · 9B
Context
131,072 tokens
License
Apache-2.0
Released
2026-08-25
Vision
No

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

§1 Characteristics by quantization band

Q2–Q3 — 2–3 bit

4.5GB weights · e.g. Q3_K_M + KV: 1.3GB at 8K · 5GB at 32K · 20GB at 128K

— Not yet rated · contributions welcome —

Q4–Q5 — 4–5 bit

5.5GB weights · e.g. Q4_K_M , mlx-4bit + KV: 1.3GB at 8K · 5GB at 32K · 20GB at 128K
Coding 4/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: one point down. Q4_K_M is 5.54GB — comfortably inside the 8GB tier.

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

    Editorial derivation: one point down; smaller models have less headroom for the token-level precision that structured calls need.

Q6–Q8 — 6–8 bit

7.5GB weights · e.g. Q6_K + KV: 1.3GB at 8K · 5GB at 32K · 20GB at 128K
Coding 5/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q6_K · aggregated · 2026-08-30

    Editorial derivation: carried from the full band.

Tool calling / agentic 6/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q6_K · aggregated · 2026-08-30

    Editorial derivation: carried from the full band. At 7.52GB this is the band to run if you have 16GB.

FP16 — full precision (fp16/bf16)

18GB weights · e.g. bf16 + KV: 1.3GB at 8K · 5GB at 32K · 20GB at 128K
Coding 5/10
data checked Aug 2026
  • vendor-model-cardSWE-bench Verified = 47.67 @ bf16 · vendor-reported · 2026-08-30

    An 8B-class model resolving nearly half of SWE-bench Verified is notable for the size; still far below the 27B-class models here. Vendor-reported.

Tool calling / agentic 6/10
data checked Aug 2026
  • vendor-model-cardBFCL v3 = 68.27 @ bf16 · vendor-reported · 2026-08-30

    The reason to pick this model at 8B: function calling is the design goal, not an afterthought. Independent BFCL listing pending as of 2026-08-30.

Factuality 4/10
data checked Aug 2026
  • vendor-model-cardMMLU = 73.8 @ bf16 · vendor-reported · 2026-08-30

    Knowledge proxy only — no hallucination benchmark lists this model yet. Nine billion parameters is a hard ceiling on stored facts.

§2 Known issues & what fixes them

ISSUE-01 mild Q2–Q3 / Q4–Q5 / Q6–Q8 / FP16 · coding, tool-calling

Released 2026-08-25. Every score here is IBM's own — including the BFCL v3 number, which is the model's main selling point and comes from a benchmark whose public leaderboard has not yet listed it.

No fix
No known fix — choose a different model

Not a defect — a caveat. Revisit when independent numbers land.

Evidence · 1 source · anecdotal
ISSUE-02 moderate Q2–Q3 / Q4–Q5 / Q6–Q8 · instruction-following, tool-calling

Granite 4 chat-template auto-detection has failed in llama.cpp, and the Granite 4 line also has an open report of quantized builds failing at very large contexts. A mis-detected template degrades tool calling quietly — the model looks dumber than it is.

Fix
Update your runtime [strong]

Pass the Granite chat template explicitly instead of relying on auto-detection, and keep llama.cpp current.

Evidence · 2 sources · community-consensus
ISSUE-03 moderate Q2–Q3 / Q4–Q5 / Q6–Q8 / FP16 · factuality

MMLU 73.8 against the 30B sibling's 80.2. Nine billion parameters caps how much world knowledge it can hold, and no hallucination benchmark lists it, so the factuality score is a knowledge proxy rather than a measurement.

Fix
RAG (retrieval) [strong]

Ground it in retrieved documents — this model is built for enterprise RAG and tool use, which is exactly the workload that avoids the ceiling.

Evidence · 1 source · anecdotal