Modelpublic

Rewnozom/Rewnozom-GGUF

sourceHugging Faceapache-2.0updated 20d agoView on Hugging Face

1likes25kdownloads

Model Card

Rewnozom-GGUF

GGUF conversion of Rewnozom/Rewnozom, derived from Qwen/Qwen2.5-7B-Instruct-1M.

—Original model: Rewnozom/Rewnozom
—Converted model: Rewnozom/Rewnozom-GGUF
—Format: GGUF
—Runtime targets: llama.cpp, Ollama, LM Studio, Jan, and other GGUF loaders

Personal recommendation: start with the two hybrid MX profiles, MX-k_quants and MX-legacy-quants, because they keep the sensitive tensors higher while leaving the rest on a compact base quant.

Available Quantizations

Only quantizations with an actual .gguf file in this workspace are listed.

Quant	File	Approx size
`F32`	`F32/Rewnozom-1M-LR.F32.gguf`	28.38 GB
`F16`	`F16/Rewnozom-1M-LR.F16.gguf`	14.19 GB
`BF16`	`BF16/Rewnozom-1M-LR.BF16.gguf`	14.19 GB
`MX-k_quants`	`MX-k_quants/Rewnozom-1M-LR.Q4_K_M.k_quants_hybrid_q4km_q5_q6.gguf`	4.62 GB
`MX-legacy-quants`	`MX-legacy-quants/Rewnozom-1M-LR.Q4_0.legacy_hybrid_q40_q50_q80.gguf`	5.05 GB
`MXFP4_MOE`	`extra_current_llama_quants/MXFP4_MOE/Rewnozom-1M-LR.MXFP4_MOE.gguf`	7.54 GB
`Q1_0`	`extra_current_llama_quants/Q1_0/Rewnozom-1M-LR.Q1_0.gguf`	1.35 GB
`Q2_0`	`extra_current_llama_quants/Q2_0/Rewnozom-1M-LR.Q2_0.gguf`	2.42 GB
`Q2_K`	`k_quants/Q2_K/Rewnozom-1M-LR.Q2_K.gguf`	2.81 GB
`Q3_K_L`	`k_quants/Q3_K_L/Rewnozom-1M-LR.Q3_K_L.gguf`	3.81 GB
`Q3_K_M`	`k_quants/Q3_K_M/Rewnozom-1M-LR.Q3_K_M.gguf`	3.55 GB
`Q3_K_S`	`k_quants/Q3_K_S/Rewnozom-1M-LR.Q3_K_S.gguf`	3.25 GB
`Q4_K_M`	`k_quants/Q4_K_M/Rewnozom-1M-LR.Q4_K_M.gguf`	4.36 GB
`Q4_K_S`	`k_quants/Q4_K_S/Rewnozom-1M-LR.Q4_K_S.gguf`	4.15 GB
`Q5_K_M`	`k_quants/Q5_K_M/Rewnozom-1M-LR.Q5_K_M.gguf`	5.07 GB
`Q5_K_S`	`k_quants/Q5_K_S/Rewnozom-1M-LR.Q5_K_S.gguf`	4.95 GB
`Q6_K`	`k_quants/Q6_K/Rewnozom-1M-LR.Q6_K.gguf`	5.82 GB
`Q4_0`	`legacy_quants/Q4_0/Rewnozom-1M-LR.Q4_0.gguf`	4.13 GB
`Q4_1`	`legacy_quants/Q4_1/Rewnozom-1M-LR.Q4_1.gguf`	4.54 GB
`Q5_0`	`legacy_quants/Q5_0/Rewnozom-1M-LR.Q5_0.gguf`	4.95 GB
`Q5_1`	`legacy_quants/Q5_1/Rewnozom-1M-LR.Q5_1.gguf`	5.36 GB
`Q8_0`	`legacy_quants/Q8_0/Rewnozom-1M-LR.Q8_0.gguf`	7.54 GB
`TQ1_0`	`t_quants/TQ1_0/Rewnozom-1M-LR.TQ1_0.gguf`	1.99 GB
`TQ2_0`	`t_quants/TQ2_0/Rewnozom-1M-LR.TQ2_0.gguf`	2.28 GB

Quantization Groups

Group	Purpose
`F32`	Full float32 GGUF reference conversion. Highest precision, largest file.
`F16`	Float16 GGUF reference conversion. Good baseline for further quantization.
`BF16`	BFloat16 GGUF reference conversion for runtimes that prefer BF16.
`extra_current_llama_quants`	Additional current llama.cpp-compatible quant types.
`k_quants`	K-quant family, usually the best default family for local use.
`legacy_quants`	Older GGUF quant family for compatibility and comparison.
`t_quants`	Ternary/low-bit quant family for very small local deployments.
`ollama_modelfiles`	Generated Ollama Modelfiles using the base system prompt from `sp.md`.

Recommended Starting Points

Use case	Quant
Personal recommendation, K-family hybrid	`MX-k_quants`
Personal recommendation, legacy hybrid	`MX-legacy-quants`
Best quality among compact K-quants	`Q6_K`
Balanced default	`Q4_K_M`
Smaller memory footprint	`Q3_K_M` or `Q3_K_S`
Very small local test	`Q2_K`, `Q2_0`, `TQ2_0`, or `Q1_0`
Legacy compatibility check	`Q4_0`, `Q4_1`, `Q5_0`, `Q5_1`, `Q8_0`

Actual quality and speed depend on runtime, CPU/GPU offload, context length, and prompt workload. Validate the target quant against your real tasks before using it as a default.

llama.cpp

Run directly from a local GGUF file:

bash

llama-cli -m k_quants/Q4_K_M/Rewnozom-1M-LR.Q4_K_M.gguf \
  -p "Review this implementation plan for missing constraints."

Start an OpenAI-compatible local server:

bash

llama-server -m k_quants/Q4_K_M/Rewnozom-1M-LR.Q4_K_M.gguf

Dataset Context

The model is associated with a synthetic reasoning/control-plane dataset family covering:

—boolean CSP logic
—branch-dependent task DAGs
—ordering and plan repair
—multi-hop forward inference
—request decomposition
—context relevance
—deterministic state transitions
—memory lifecycle
—retrieval/navigation policy
—executor routing
—tool execution
—permissions
—multi-agent orchestration
—result validation
—retry/escalation
—composite execution kernel behavior

The dataset design uses deterministic formal worlds, double oracle checks, structural dedupe, and machine-verifiable answers rather than synthetic prose chain-of-thought.

Limitations

—GGUF quantization changes numerical behavior compared with the source model.
—Lower-bit quants trade quality for memory and speed.
—Permission enforcement, destructive actions, and state mutation should remain controlled by deterministic application logic.

Attribution

This conversion is based on Rewnozom/Rewnozom, which is derived from Qwen/Qwen2.5-7B-Instruct-1M and follows the Apache 2.0 license.

page:

Base model: https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-1M

Rewnozom/Rewnozom: https://huggingface.co/Rewnozom/Rewnozom

Rewnozom/Rewnozom-GGUF: https://huggingface.co/Rewnozom/Rewnozom-GGUF

Ollama: https://ollama.com/tobraa92/Rewnozom

Portfolio: https://tobiasraanaes.se/