minjaechoi

Midm-2.0-Mini-Instruct-W4A16

Quantization method

Table

Method	GPTQ (`GPTQModifier`, LLM Compressor)
Scheme	W4A16 (4-bit weights, 16-bit activations)
Weight dtype	INT4, symmetric
Group size	128
Quantized targets	`Linear` (all layers except `lm_head`)
Activation ordering	static
Dampening frac	0.01
Output format	`compressed-tensors` (`pack-quantized`)

Recipe used (recipe.yaml, included in this repo):

yaml
default_stage:
  default_modifiers:
    GPTQModifier:
      targets: [Linear]
      ignore: [lm_head]
      scheme: W4A16
      block_size: 128
      dampening_frac: 0.01
      actorder: static
      requires_calibration_data: true

Calibration set

256 packed sequences × 2048 tokens = 524,288 calibration tokens, sampled (seed=42) from a mixed Korean/English instruction + function-calling corpus, targeting the following source composition:

Table with columns: Source, Target %, Actual % (this model), Tokens (this model)
Source	Target %	Actual % (this model)	Tokens (this model)
KRX-Data/Won-Instruct	35%	34.19%	179,256
heegyu/glaive-function-calling-v2-ko	30%	29.80%	156,212
NousResearch/hermes-function-calling-v1 (json-mode-agentic.json)	10%

Raw pool before filtering/dedup: 49,083 examples (49,017 after removing 65 duplicates)
Quota-sampled for calibration: 555 examples → 536 packed into the final 256×2048 blocks
Tool-calling trajectories in the pool: 8,897 (2,819 multi-tool)
Contamination check: gsm8k-ko: train split only used; no BFCL/AgentDojo/tau-bench/SafeCommit-eval sources included

safecommit_synth is unreleased in-house synthetic data from the SafeCommit project, not a public HF dataset.

How to run

vLLM (recommended — required for the `compressed-tensors` W4A16 kernels used here)

bash
vllm serve minjaechoi/Midm-2.0-Mini-Instruct-W4A16 --served-model-name midm2-mini

bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "midm2-mini",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Transformers (requires the `compressed-tensors` package for int4 dequant kernels)

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "minjaechoi/Midm-2.0-Mini-Instruct-W4A16"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

msgs = [{"role": "user", "content": "Hello!"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256)
print(tok.decode(out[0], skip_special_tokens=True))

Benchmarks

Evaluated on this W4A16 checkpoint (not compared against an FP16 baseline run in this project yet):

AgentDojo (banking suite, utility-only, 16 tasks): 5/16 tasks passed (31.2%) — utility-only scoring, no prompt-injection attacks included.
BFCL v4: attempted, but the evaluation harness used had a request-routing bug (--local-model-path was not substituted into the API request, so every call 404'd against the vLLM server) — the resulting scores are invalid and are intentionally not published here. Will be updated after a corrected re-run.

Files

model.safetensors — quantized weights (compressed-tensors pack-quantized format)
config.json — includes the quantization_config (compressed-tensors) needed by vLLM/transformers to load this checkpoint
recipe.yaml — the exact LLM Compressor recipe used to produce this checkpoint
tokenizer.json, tokenizer_config.json, chat_template.jinja — tokenizer/chat template, copied unmodified from the base model
LICENSE — base model license, included per its terms

License

This checkpoint is a derivative of K-intelligence/Midm-2.0-Mini-Instruct and is distributed under the same license (mit, see LICENSE in this repo). No additional restrictions are added beyond the base model's license.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

minjaechoi

Model Tree

Base

K-intelligence/Midm-2.0-Mini-Instruct

Quantized

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

Quantization method

Table

Method	GPTQ (`GPTQModifier`, LLM Compressor)
Scheme	W4A16 (4-bit weights, 16-bit activations)
Weight dtype	INT4, symmetric
Group size	128
Quantized targets	`Linear` (all layers except `lm_head`)
Activation ordering	static
Dampening frac	0.01
Output format	`compressed-tensors` (`pack-quantized`)

Recipe used (recipe.yaml, included in this repo):

yaml
default_stage:
  default_modifiers:
    GPTQModifier:
      targets: [Linear]
      ignore: [lm_head]
      scheme: W4A16
      block_size: 128
      dampening_frac: 0.01
      actorder: static
      requires_calibration_data: true

Calibration set

256 packed sequences × 2048 tokens = 524,288 calibration tokens, sampled (seed=42) from a mixed Korean/English instruction + function-calling corpus, targeting the following source composition:

Table with columns: Source, Target %, Actual % (this model), Tokens (this model)
Source	Target %	Actual % (this model)	Tokens (this model)
KRX-Data/Won-Instruct	35%	34.19%	179,256
heegyu/glaive-function-calling-v2-ko	30%	29.80%	156,212
NousResearch/hermes-function-calling-v1 (json-mode-agentic.json)	10%

Raw pool before filtering/dedup: 49,083 examples (49,017 after removing 65 duplicates)
Quota-sampled for calibration: 555 examples → 536 packed into the final 256×2048 blocks
Tool-calling trajectories in the pool: 8,897 (2,819 multi-tool)
Contamination check: gsm8k-ko: train split only used; no BFCL/AgentDojo/tau-bench/SafeCommit-eval sources included

safecommit_synth is unreleased in-house synthetic data from the SafeCommit project, not a public HF dataset.

How to run

vLLM (recommended — required for the `compressed-tensors` W4A16 kernels used here)

bash
vllm serve minjaechoi/Midm-2.0-Mini-Instruct-W4A16 --served-model-name midm2-mini

bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "midm2-mini",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Transformers (requires the `compressed-tensors` package for int4 dequant kernels)

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "minjaechoi/Midm-2.0-Mini-Instruct-W4A16"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

msgs = [{"role": "user", "content": "Hello!"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256)
print(tok.decode(out[0], skip_special_tokens=True))

Benchmarks

Evaluated on this W4A16 checkpoint (not compared against an FP16 baseline run in this project yet):

AgentDojo (banking suite, utility-only, 16 tasks): 5/16 tasks passed (31.2%) — utility-only scoring, no prompt-injection attacks included.
BFCL v4: attempted, but the evaluation harness used had a request-routing bug (--local-model-path was not substituted into the API request, so every call 404'd against the vLLM server) — the resulting scores are invalid and are intentionally not published here. Will be updated after a corrected re-run.

Files

model.safetensors — quantized weights (compressed-tensors pack-quantized format)
config.json — includes the quantization_config (compressed-tensors) needed by vLLM/transformers to load this checkpoint
recipe.yaml — the exact LLM Compressor recipe used to produce this checkpoint
tokenizer.json, tokenizer_config.json, chat_template.jinja — tokenizer/chat template, copied unmodified from the base model
LICENSE — base model license, included per its terms

Midm-2.0-Mini-Instruct-W4A16

README

Quantization method

Calibration set

How to run

vLLM (recommended — required for the `compressed-tensors` W4A16 kernels used here)

Transformers (requires the `compressed-tensors` package for int4 dequant kernels)

Benchmarks

Files

License

Explore FriendliAI today

README

Quantization method

Calibration set

How to run

vLLM (recommended — required for the `compressed-tensors` W4A16 kernels used here)

Transformers (requires the `compressed-tensors` package for int4 dequant kernels)

Benchmarks

Files

License

Midm-2.0-Mini-Instruct-W4A16

README

Quantization method

Calibration set

How to run

vLLM (recommended — required for the compressed-tensors W4A16 kernels used here)

Transformers (requires the compressed-tensors package for int4 dequant kernels)

Benchmarks

Files

License

Explore FriendliAI today

README

Quantization method

Calibration set

How to run

vLLM (recommended — required for the compressed-tensors W4A16 kernels used here)

Transformers (requires the compressed-tensors package for int4 dequant kernels)

Benchmarks

Files

License

vLLM (recommended — required for the `compressed-tensors` W4A16 kernels used here)

Transformers (requires the `compressed-tensors` package for int4 dequant kernels)

vLLM (recommended — required for the `compressed-tensors` W4A16 kernels used here)

Transformers (requires the `compressed-tensors` package for int4 dequant kernels)