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

Model datasheet · Phi

Phi-4-reasoning-vision

Microsoft's selective-reasoning vision model — it decides when to think. Catalogued for completeness with a hard local caveat: llama.cpp cannot run its vision tower, so local GGUFs are text-only; the headline capability needs the full weights via transformers.

Vendor
Microsoft
Architecture
Dense · 15B
Context
16,384 tokens
License
MIT
Released
2026-03-04
Vision
Yes

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

§1 Characteristics by quantization band

Q2–Q3 — 2–3 bit

7GB weights · e.g. Q3_K_M + KV: 1.6GB at 8K · 3.1GB at 32K · 3.1GB at 128K

— Not yet rated · contributions welcome —

Q4–Q5 — 4–5 bit

9GB weights · e.g. Q4_K_M , mlx-4bit + KV: 1.6GB at 8K · 3.1GB at 32K · 3.1GB at 128K

— Not yet rated · contributions welcome —

Q6–Q8 — 6–8 bit

12GB weights · e.g. Q6_K , Q8_0 + KV: 1.6GB at 8K · 3.1GB at 32K · 3.1GB at 128K

— Not yet rated · contributions welcome —

FP16 — full precision (fp16/bf16)

29GB weights · e.g. fp16 + KV: 1.6GB at 8K · 3.1GB at 32K · 3.1GB at 128K

— Not yet rated · contributions welcome —

§2 Known issues & what fixes them

ISSUE-01 severe Q2–Q3 / Q4–Q5 / Q6–Q8 / FP16 · instruction-following

Its reason-to-exist doesn't work in llama.cpp: there is no support for the phi4-siglip vision tower, so every GGUF is a text-only backbone (which is why our text dimensions sit unrated). An mlx-community 4-bit build exists and may be the practical local vision path on Apple Silicon via mlx-vlm — unverified by us.

Fix
Update your runtime [partial]

On a Mac, try the mlx-community 4-bit build with mlx-vlm. Elsewhere, watch llama.cpp for phi4-siglip support; until then vision needs full weights via transformers.

No fix
No known fix — choose a different model

For local vision today, use Gemma 3, Mistral Small 4, or Qwen3.8 instead.

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

Inherits Phi-4's 16K window — cramped for 2026, especially for the multi-image reasoning it was built for.

Fix
RAG (retrieval) [partial]

Chunked retrieval keeps requests inside 16K.

Evidence · 1 source · anecdotal