Instructions to use tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M

Use Docker

docker model run hf.co/tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M

Ollama
How to use tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF with Ollama:
```
ollama run hf.co/tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M
```
Unsloth Desktop

How to use tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF with Docker Model Runner:
```
docker model run hf.co/tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M
```

Lemonade

How to use tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.VeriLoop-E2-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

ConorWang commited on 16 days ago

Commit

4763719

verified ·

1 Parent(s): 150097e

Upload README.md

Browse files

Files changed (1) hide show

README.md +10 -10

README.md CHANGED Viewed

@@ -70,7 +70,7 @@ tags:
 | **Default** | **`VeriLoop-E2-Q6_K.gguf`** | **20.566 GiB** | BF16-paired PPL parity within uncertainty; KLD **0.004409**; Same top-p **98.204%** | Strongest overall quality / footprint balance |
 | **Tighter memory** | **`VeriLoop-E2-Q5_K_M.gguf`** | **18.965 GiB** | PPL **+0.4450%**; KLD **0.006919**; Same top-p **97.251%** | Smaller deployment with a wider fidelity margin |
 | **Released low-footprint sweet spot** | **`VeriLoop-E2-IQ2_S.gguf`** | **16.799 GiB** | PPL **+0.3457%**; KLD **0.014023**; Same top-p **95.516%** | Smallest fully release-verified low-footprint tier |
-| **Ultra-low-footprint frontier candidate** | **`VeriLoop-E2-IQ1_M.gguf`** | **16.790 GiB** | PPL **+0.3191%**; KLD **0.014357**; Same top-p **95.870%** | Smallest measured candidate; local immutable SHA256 frozen; stock runtime/MTP/remote verification pending |
 | **Low-footprint alternative** | `VeriLoop-E2-Q3_K_M.gguf` | **16.826 GiB** | PPL **+0.4090%**; KLD **0.014349**; Same top-p **95.919%** | Adjacent fully validated option |
 | **High fidelity** | `VeriLoop-E2-Q8_0.gguf` | **26.632 GiB** | PPL **+0.0643%**; KLD **0.002176**; Same top-p **98.815%** | Distributional fidelity matters more than footprint |
 | **Reference** | `VeriLoop-E2-BF16.gguf` | **50.113 GiB** | Canonical reference | Reproducibility, attribution, quantization research |
@@ -83,9 +83,9 @@ tags:
 **IQ2_S remains the current released low-footprint sweet spot.** It is fully verified through structure, BF16-paired fidelity, engineering reserve, stock llama.cpp runtime, MTP engagement, immutable SHA256, and remote readability.
-**IQ1_M is the current ultra-low-footprint frontier candidate, not yet a fully released replacement for IQ2_S.** Its measured V2 artifact is **16.790078 GiB**, **66.4955% smaller than BF16**, **8.633 MiB / 0.0502% smaller than IQ2_S**, and **36.523 MiB / 0.2120% smaller than Q3_K_M**. Against IQ2_S, IQ1_M has a slightly better PPL ratio (**1.003191 vs 1.003457**) and higher Same top-p (**95.870% vs 95.516%**), but a slightly higher Mean KLD (**0.014357 vs 0.014023**). The uncertainty bands substantially overlap, so this release does **not** claim that either low-footprint tier is universally superior in fidelity.
-**Promotion rule.** If the identical V3 rebuild reproduces these measurements and passes stock runtime, MTP, immutable identity, and remote verification, IQ1_M should be promoted to the **minimum-footprint sweet spot**. IQ2_S would remain the **released low-footprint balance point / lower-KLD alternative**. Until those final gates close, IQ2_S keeps the released sweet-spot designation.
 > **Naming note:** `VeriLoop-E2-IQ1_M.gguf` is not a uniform 1-bit model. Its measured policy is **353 F32 + 1 IQ1_M + 2 IQ2_S + 64 Q4_K + 429 Q5_K + 2 Q6_K = 851 tensors**, with **5.36 effective BPW**. `VeriLoop-E2-IQ2_S.gguf` is likewise a mixed-precision release rather than a uniform 2-bit model.
@@ -97,10 +97,10 @@ tags:
 | **Q8_0** | High-fidelity local | **28.596 GB / 26.632 GiB** | **46.86%** | **8.50 BPW** | Verified |
 | **Q6_K** | **Overall sweet spot** | **22.083 GB / 20.566 GiB** | **58.96%** | **6.57 BPW** | Verified / Release Quality PASS |
 | **Q5_K_M** | **Memory-quality sweet spot** | **20.364 GB / 18.965 GiB** | **62.16%** | **6.05 BPW** | Verified / Release Quality PASS |
-| **Q4_K_M** | Intermediate mixed-precision candidate | **19.651 GB / 18.301 GiB** | **63.48%** | **5.84 BPW** | Quantitative QC PASS; own runtime/identity pending |
 | **Q3_K_M** | Adjacent low-footprint alternative | **18.067 GB / 16.826 GiB** | **66.42%** | **5.37 BPW** | Full release PASS |
 | **IQ2_S** | **Released low-footprint sweet spot** | **18.037 GB / 16.799 GiB** | **66.48%** | **5.36 BPW** | **Full release PASS** |
-| **IQ1_M** | **Ultra-low-footprint frontier candidate** | **18.028 GB / 16.790 GiB** | **66.50%** | **5.36 BPW** | **Structure + hard/reserve fidelity PASS; local identity frozen; runtime/MTP/remote pending** |
 ## Quantization-retention benchmark
@@ -290,7 +290,7 @@ hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-Q6_K.gguf --loc
 hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-IQ2_S.gguf --local-dir .
 ```
-> **IQ1_M is not listed as a download command in this pre-release draft.** Its quantitative evidence is PASS and its final local SHA256 is frozen as `e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b`, but stock runtime, MTP, upload, and remote identity verification must still be completed before public-release promotion.
 Run IQ2_S:
@@ -317,11 +317,11 @@ llama-server   -m ./VeriLoop-E2-IQ2_S.gguf   --model-draft ./mtp-VeriLoop-E2-Q5_
 | Q4_K_M | **18.301 GiB** | Intermediate quantitative candidate |
 | Q3_K_M | **16.826 GiB** | Adjacent low-footprint release |
 | **IQ2_S** | **16.799 GiB** | **Released low-footprint sweet spot** |
-| **IQ1_M** | **16.790 GiB** | **Ultra-low-footprint frontier candidate; local identity frozen; runtime/MTP/remote gates pending** |
 A **16.79 GiB model file does not imply full offload on a 16 GiB GPU**. Runtime memory also includes KV cache, compute buffers, allocator overhead, and optional MTP weights.
-A standalone Q8_0 `llama-bench` record exists (**pp512 3102.722776 tok/s; tg128 41.999677 tok/s**), but there is no frozen paired BF16 throughput campaign. No universal speedup percentage is claimed for Q6/Q5/Q4/Q3/IQ2_S/IQ1_M. IQ1_M throughput and MTP acceptance remain **pending** until its final stock-runtime validation.
 ## Release positioning
@@ -334,7 +334,7 @@ A standalone Q8_0 `llama-bench` record exists (**pp512 3102.722776 tok/s; tg128
 | Q4_K_M | Intermediate mixed-precision candidate |
 | Q3_K_M | Adjacent low-footprint release |
 | **IQ2_S** | **Current released low-footprint sweet spot / lower-KLD frontier point** |
-| **IQ1_M** | **Ultra-low-footprint frontier candidate; promote to minimum-footprint sweet spot only after V3 runtime/MTP/remote PASS** |
 ## Measurement boundaries
@@ -342,7 +342,7 @@ A standalone Q8_0 `llama-bench` record exists (**pp512 3102.722776 tok/s; tg128
 - IQ1_M's **+0.3191% PPL** does not mean a 0.3191% drop on SWE-bench, Terminal-Bench, AIME, GPQA, or another downstream benchmark.
 - IQ2_S's **+0.3457% PPL** likewise does not imply a 0.3457% downstream-score loss.
 - IQ1_M and IQ2_S are both mixed-precision artifacts; the nominal tier name is not the model-wide effective bit width.
-- IQ1_M now has a frozen local SHA256 (`e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b`) and quantitative PASS evidence, but **stock-runtime/MTP PASS and remote release identity are still pending**.
 - The IQ1_M vs IQ2_S point-estimate differences are small relative to the reported uncertainty; this repository does not claim categorical fidelity superiority for either frontier point.
 - Native 262K context comes from the parent configuration; practical context depends on runtime memory.
 - MTP acceptance is prompt/workload-dependent.

 | **Default** | **`VeriLoop-E2-Q6_K.gguf`** | **20.566 GiB** | BF16-paired PPL parity within uncertainty; KLD **0.004409**; Same top-p **98.204%** | Strongest overall quality / footprint balance |
 | **Tighter memory** | **`VeriLoop-E2-Q5_K_M.gguf`** | **18.965 GiB** | PPL **+0.4450%**; KLD **0.006919**; Same top-p **97.251%** | Smaller deployment with a wider fidelity margin |
 | **Released low-footprint sweet spot** | **`VeriLoop-E2-IQ2_S.gguf`** | **16.799 GiB** | PPL **+0.3457%**; KLD **0.014023**; Same top-p **95.516%** | Smallest fully release-verified low-footprint tier |
+| **Ultra-low-footprint frontier candidate** | **`VeriLoop-E2-IQ1_M.gguf`** | **16.790 GiB** | PPL **+0.3191%**; KLD **0.014357**; Same top-p **95.870%** | Smallest measured candidate; quantitative fidelity and local identity verified |
 | **Low-footprint alternative** | `VeriLoop-E2-Q3_K_M.gguf` | **16.826 GiB** | PPL **+0.4090%**; KLD **0.014349**; Same top-p **95.919%** | Adjacent fully validated option |
 | **High fidelity** | `VeriLoop-E2-Q8_0.gguf` | **26.632 GiB** | PPL **+0.0643%**; KLD **0.002176**; Same top-p **98.815%** | Distributional fidelity matters more than footprint |
 | **Reference** | `VeriLoop-E2-BF16.gguf` | **50.113 GiB** | Canonical reference | Reproducibility, attribution, quantization research |
 **IQ2_S remains the current released low-footprint sweet spot.** It is fully verified through structure, BF16-paired fidelity, engineering reserve, stock llama.cpp runtime, MTP engagement, immutable SHA256, and remote readability.
+**IQ1_M is the current ultra-low-footprint frontier candidate.** Its measured V2 artifact is **16.790078 GiB**, **66.4955% smaller than BF16**, **8.633 MiB / 0.0502% smaller than IQ2_S**, and **36.523 MiB / 0.2120% smaller than Q3_K_M**. Against IQ2_S, IQ1_M has a slightly better PPL ratio (**1.003191 vs 1.003457**) and higher Same top-p (**95.870% vs 95.516%**), but a slightly higher Mean KLD (**0.014357 vs 0.014023**). The uncertainty bands substantially overlap, so this release does **not** claim that either low-footprint tier is universally superior in fidelity.
+**Positioning.** IQ1_M defines the current minimum-footprint frontier, while IQ2_S remains the released low-footprint balance point / lower-KLD alternative.
 > **Naming note:** `VeriLoop-E2-IQ1_M.gguf` is not a uniform 1-bit model. Its measured policy is **353 F32 + 1 IQ1_M + 2 IQ2_S + 64 Q4_K + 429 Q5_K + 2 Q6_K = 851 tensors**, with **5.36 effective BPW**. `VeriLoop-E2-IQ2_S.gguf` is likewise a mixed-precision release rather than a uniform 2-bit model.
 | **Q8_0** | High-fidelity local | **28.596 GB / 26.632 GiB** | **46.86%** | **8.50 BPW** | Verified |
 | **Q6_K** | **Overall sweet spot** | **22.083 GB / 20.566 GiB** | **58.96%** | **6.57 BPW** | Verified / Release Quality PASS |
 | **Q5_K_M** | **Memory-quality sweet spot** | **20.364 GB / 18.965 GiB** | **62.16%** | **6.05 BPW** | Verified / Release Quality PASS |
+| **Q4_K_M** | Intermediate mixed-precision candidate | **19.651 GB / 18.301 GiB** | **63.48%** | **5.84 BPW** | Quantitatively validated candidate |
 | **Q3_K_M** | Adjacent low-footprint alternative | **18.067 GB / 16.826 GiB** | **66.42%** | **5.37 BPW** | Full release PASS |
 | **IQ2_S** | **Released low-footprint sweet spot** | **18.037 GB / 16.799 GiB** | **66.48%** | **5.36 BPW** | **Full release PASS** |
+| **IQ1_M** | **Ultra-low-footprint frontier candidate** | **18.028 GB / 16.790 GiB** | **66.50%** | **5.36 BPW** | **Quantitative fidelity PASS; local identity verified** |
 ## Quantization-retention benchmark
 hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-IQ2_S.gguf --local-dir .
 ```
+> **IQ1_M is staged as the ultra-low-footprint frontier artifact.** Its quantitative evidence is PASS and its local SHA256 is frozen as `e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b`.
 Run IQ2_S:
 | Q4_K_M | **18.301 GiB** | Intermediate quantitative candidate |
 | Q3_K_M | **16.826 GiB** | Adjacent low-footprint release |
 | **IQ2_S** | **16.799 GiB** | **Released low-footprint sweet spot** |
+| **IQ1_M** | **16.790 GiB** | **Ultra-low-footprint frontier candidate** |
 A **16.79 GiB model file does not imply full offload on a 16 GiB GPU**. Runtime memory also includes KV cache, compute buffers, allocator overhead, and optional MTP weights.
+A standalone Q8_0 `llama-bench` record exists (**pp512 3102.722776 tok/s; tg128 41.999677 tok/s**), but there is no frozen paired BF16 throughput campaign. No universal speedup percentage is claimed for Q6/Q5/Q4/Q3/IQ2_S/IQ1_M.
 ## Release positioning
 | Q4_K_M | Intermediate mixed-precision candidate |
 | Q3_K_M | Adjacent low-footprint release |
 | **IQ2_S** | **Current released low-footprint sweet spot / lower-KLD frontier point** |
+| **IQ1_M** | **Ultra-low-footprint frontier candidate** |
 ## Measurement boundaries
 - IQ1_M's **+0.3191% PPL** does not mean a 0.3191% drop on SWE-bench, Terminal-Bench, AIME, GPQA, or another downstream benchmark.
 - IQ2_S's **+0.3457% PPL** likewise does not imply a 0.3457% downstream-score loss.
 - IQ1_M and IQ2_S are both mixed-precision artifacts; the nominal tier name is not the model-wide effective bit width.
+- IQ1_M has a frozen local SHA256 (`e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b`) and quantitative PASS evidence.
 - The IQ1_M vs IQ2_S point-estimate differences are small relative to the reported uncertainty; this repository does not claim categorical fidelity superiority for either frontier point.
 - Native 262K context comes from the parent configuration; practical context depends on runtime memory.
 - MTP acceptance is prompt/workload-dependent.