Jun 19, 2026

LLM Usage notes

   ┌──────────────────────────────────────────────────────────────────────────┐
   │      ▓▓▓▒▒▒ PROVENANCE NODE 001-1 // ZONE: HMN_REMIXABLE ▒▒▒▓▓▓          │
   ├──────────────────────────────────────────────────────────────────────────┤
   │ SRC_AUTH : HMN                    │ SIG_STAT : UNSIGNED                  │
   │ AUTH_ID  : NICK PORCINO           │ TYPE     : SURVEY                    │
   │ DATE     : 20260619               │ TYPE     : SURVEY                    │
   ├──────────────────────────────────────────────────────────────────────────┤
   │ [ RIGHTS & PERMISSIONS MATRIX ]                                          │
   │ INDEX_ALLOW : YES                 │ CORP_TRAIN: NO                       │
   │ DERIV_ALLOW : YES                 │ GOV_SPDX  : ResponsibleSrc           │
   └──────────────────────────────────────────────────────────────────────────┘

Introduction

The following is a brief accounting of models found to be useful, and some characterizing notes based on experience using those models. Scripts to download and test most of these models may be found here:

https://codeberg.org/meshula/LabLlama/src/branch/dev/agentic/scripts

Not all models available via those scripts are useful enough to warrant further notes, and won’t be found below. If sufficient interest in this list develops, future work will include proper benchmarking via a system such as terminal bench, or the intelligence per watt benchmark listed in the notes.

QWEN

Qwen3.6 27B NVFP4

TAG	VALUE
MODEL	Qwen3.6 27B NVFP4
DATE	20260618
PROC	DGX SPARK 128GB
SRV	UNKNOWN
PROMPT/GEN	N/A
THINKING	EXCELLENT
COHERENCE	EXCELLENT
PLANNING	EXCELLENT
CODING	UNTESTED
SYCOPHANCY	NOT OBNOXIOUS

Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16-mlx-fp16

TAG	VALUE
MODEL	Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16-mlx-fp16
DATE	20260610
PROC	M5 MAX 128GB
SRV	OMLX
PROMPT/GEN	N/10
THINKING	EXCELLENT
COHERENCE	EXCELLENT
PLANNING	EXCELLENT
CODING	GOOD, SOME HALLUCINATION
SYCOPHANCY	VERY LOW

Qwen3-Coder-Next-Q6_K

TAG	VALUE
MODEL	Qwen3-Coder-Next-Q6_K
DATE	20260524
PROC	M3 MAX 128GB
SRV	LLAMA-SERVE
PROMPT/GEN	30/10
THINKING	N/A
COHERENCE	GOOD
PLANNING	GOOD
CODING	GOOD, HALLUCINATION
SYCOPHANCY	LOW

Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF

TAG	VALUE
MODEL	Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF
DATE	20260524
PROC	M3 MAX 128GB
SRV	LLAMA-SERVE
PROMPT/GEN	30/10
THINKING	EXCELLENT
COHERENCE	EXCELLENT
PLANNING	EXCELLENT
CODING	GOOD
SYCOPHANCY	LOW

Gemma

unsloth/gemma-4-26B-A4B-it-qat-GGUF:UD-Q4_K_XL

TAG	VALUE
MODEL	unsloth/gemma-4-26B-A4B-it-qat-GGUF:UD-Q4_K_XL
DATE	20260618
PROC	M1 ULTRA 64GB
SRV	LLAMA-SERVE —spec-type draft-mtp —spec-draft-n-max 2
PROMPT/GEN	800/70
THINKING	EXCELLENT
COHERENCE	EXCELLENT
PLANNING	UNTESTED
CODING	UNTESTED
SYCOPHANCY	VERY LOW

gemma-4-26B-A4B-it-unsloth-mlx-oQ4-fp16

TAG	VALUE
MODEL	gemma-4-26B-A4B-it-unsloth-mlx-oQ4-fp16
DATE	20260610
PROC	M5 MAX 128GB
SRV	OMLX
PROMPT/GEN	N/30
THINKING	EXCELLENT
COHERENCE	EXCELLENT
PLANNING	EXCELLENT
CODING	GOOD, SOME HALLUCINATION
SYCOPHANCY	MODERATE

GPT-OSS

gpt-oss-120b-mxfp4

TAG	VALUE
MODEL	gpt-oss-120b-mxfp4
DATE	20260524
PROC	M3 MAX 128GB
SRV	LLAMA-SERVE
PROMPT/GEN	30/10
THINKING	EXCELLENT
COHERENCE	EXCELLENT
PLANNING	EXCELLENT
CODING	GOOD
SYCOPHANCY	MODERATE

OTHER

Codestral-22B-v0.1.Q4_K_M

TAG	VALUE
MODEL	Codestral-22B-v0.1.Q4_K_M
DATE	20260524
PROC	M3 MAX 128GB
SRV	LLAMA-SERVE
PROMPT/GEN	N/30
THINKING	N/A
COHERENCE	GOOD
PLANNING	MODERATE
CODING	MODERATE, HALLUCINATION
SYCOPHANCY	LOW

Mixtral-8x7B-Instruct-v0.1.Q4_K_M

TAG	VALUE
MODEL	Mixtral-8x7B-Instruct-v0.1.Q4_K_M
DATE	20260524
PROC	M3 MAX 128GB
SRV	LLAMA-SERVE
PROMPT/GEN	N/30
THINKING	N/A
COHERENCE	POOR
PLANNING	N/A
CODING	HIGH HALLUCINATION
SYCOPHANCY	MARCH HARE