Voltron-1 Mini
Voltron-1 Mini (Akahsizrr/voltron-1-mini) is a 1.73B-parameter language model built on the
LFM3-1.7B architecture by
Liquid AI โ a hybrid model combining gated short convolutions with
sliding-window and global attention. This checkpoint keeps the same
architecture, tokenizer, and chat format as the base model, with weights
optimized for stronger coding, mathematical reasoning, and instruction
following.
- Parameters: 1.73B
- Layers: 24 (hybrid short-conv + attention)
- Hidden size: 2,048
- Context length: 262,144 tokens
- Vocabulary: shared LFM3 tokenizer (128K)
- Weights: BF16 safetensors
- License: LFM Open License (same as
LiquidAI/LFM3-1.7B)
Model summary
Voltron-1 Mini is intended as a drop-in upgrade to
LiquidAI/LFM3-1.7B for chat, code generation, and reasoning workloads.
It is particularly strong at competitive-programming-style tasks and
multi-step math reasoning relative to its size, while retaining the base
model's instruction-following behavior.
Benchmark results
All scores below were measured by us with thinking enabled on a
single H100-class GPU using vLLM (BF16, temperature 0.6, top-p 0.95,
seeded sampling). LiveCodeBench uses the test6 split (175 problems,
pass@k over 5 samples); AIME uses AIME-2026 (30 problems ร 4 samples);
HumanEval uses the full 164-problem set; IFEval uses the official
instruction classes (strict and loose); GPQA uses the 198-question
Diamond split.
Head-to-head vs base LFM3-1.7B (same protocol)
| Benchmark | LFM3-1.7B (base) | Voltron-1 Mini | ฮ |
|---|---|---|---|
| LiveCodeBench pass@1 | 0.2011 | 0.2434 | +4.2 pp |
| LiveCodeBench pass@5 | 0.2571 | 0.3257 | +6.9 pp |
| HumanEval pass@1 | 0.6402 | 0.6707 | +3.1 pp |
| AIME2026 accuracy | 0.408 | 0.442 | +3.4 pp |
| AIME2026 pass@any | 0.500 | 0.567 | +6.7 pp |
| IFEval strict | 0.7246 | 0.7209 | โ0.4 pp |
| IFEval loose | 0.7782 | 0.7560 | โ2.2 pp |
| GPQA Diamond | 0.4293 | 0.4646 | +3.5 pp |
Published reference scores (different harnesses โ not directly comparable)
| Benchmark | LFM2.5-2.6B | Qwen3-4B-Instruct-2507 |
|---|---|---|
| LiveCodeBench v6 | 59.4 (vendor) | 35.1 (vendor, 25.02โ25.05 window) |
| HumanEval / HumanEval+ | 83.1 HumanEval+ (verified selection) | 77.6 |
| AIME 2025 | 51.9 (vendor) | 47.4 (vendor) |
| IFEval | 95.4 (vendor) | 83.4 |
| GPQA Diamond | 55.8 (Artificial Analysis) | 62.0 |
Note: LFM2.5-2.6B numbers come from vendor and third-party published results using different prompts, sampling, and harnesses. They are shown for rough scale only; no public results exist for these models on our exact protocol.
Voltron-1 Mini AIME scores were measured on the pre-DPO checkpoint; the final preference update was small and subsequent results stayed within noise.
Evaluation artifacts
Raw per-problem generations and scorer outputs are in eval/.
Key files: eval/lcb_dpo_v10.json (LCB), eval/he_dpo_v10.json
(HumanEval), eval/aime_gn_v10.json (AIME), and eval/r6/ (base-model
baselines plus IFEval/GPQA for base and V10).
Intended use
- General-purpose chat and assistant workloads
- Code generation and competitive-programming-style problem solving
- Mathematical and multi-step reasoning tasks
- Instruction-following tasks with verifiable constraints
Limitations
- Size: at 1.73B parameters, this is a small model. Expect weaker performance than 7B+ models on open-ended knowledge work, long-horizon agentic tasks, and niche domains.
- Benchmark variance: LiveCodeBench
test6pass@1 has a standard error of roughly ยฑ2โ3 pp at n=175ร5; small differences between checkpoints may be noise. - Instruction following: IFEval loose score is ~2 pp below base โ heavy reasoning-oriented tuning trades a little formatting strictness.
- Languages: primarily English; multilingual ability mirrors the base model's training distribution.
- Thinking traces: the model emits
<think>blocks when reasoning; disable via chat template flags for latency-sensitive use.
Usage
vLLM (recommended)
pip install vllm
vllm serve Akahsizrr/voltron-1-mini
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
resp = client.chat.completions.create(
model="Akahsizrr/voltron-1-mini",
messages=[{"role": "user", "content": "Write a Python function that returns the nth Fibonacci number."}],
temperature=0.6, top_p=0.95, max_tokens=2048,
)
print(resp.choices[0].message.content)
Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Akahsizrr/voltron-1-mini"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", dtype="bfloat16")
messages = [{"role": "user", "content": "Explain quicksort in one paragraph."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True,
return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, do_sample=True,
temperature=0.6, top_p=0.95)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
GGUF / llama.cpp
The checkpoint can be converted to GGUF with llama.cpp's convert_hf_to_gguf.py
for CPU and edge deployment; LFM3 support in llama.cpp is available in
recent releases.
Training & provenance
This model is derived from LiquidAI/LFM3-1.7B and is distributed for
research and production use under the LFM Open License. See the
LiquidAI model card for base
model details, architecture documentation, and license terms.
- Downloads last month
- 55
Model tree for Akahsizrr/voltron-1-mini
Base model
LiquidAI/LFM3-1.7B-Base