MAGI 1.3-1.7B

Paper: Where the Bits Go (Sun, 2026) ยท Recipe: GEHIRNS/magi ยท True params 1728801536 - GGUF base IQ2_XXS with 167-type mixed override ยท DOI 10.5281/zenodo.23630199

Introduction

magi-1.3-1.7B is LFM3-1.7B in 863 MiB at 32K context, highest capability density we have measured: rho=2.207 on full 2313-task bench, vs 1.305 for magi-1.2-2B, 1.149 for magi-1.1-2B and 0.385 for Qwen3.5-2B BF16, all same settings. Scores 72.45% on 2313 tasks in 0.84 GiB, 5.36 points higher than last gen 2B in 400 MiB less memory, thinks 41% less than magi-1.2-2B and 57% less than magi-1.1-2B.

LFM3-1.7B base is LiquidAI hybrid: 24 layers = 13 conv shortconv.conv q8_0, 6 full_attn 2,6,10,14,18,22 q4_k, 5 sliding w256, hidden 2048, 16 heads 4 kv, intermediate 7680, vocab 128k, max 262k, KV 113MB @32K q4_0 28MB @8192 per slot. No layers pruned after imatrix โ€” all 24 blk intact, imatrix only quant importance.

Two steps:

  1. SFT commonsense-heavy: QLoRA r16 a32 on w1,w2,w3,q_proj,k_proj,v_proj,o_proj,gate_proj,in_proj,out_proj, 1696 traces 88% commonsense (hellaswag 1400 winogrande 600 obqa 400 sciq 400 arc 400 mmlu 300 gsm8k 100), loss 0.668->0.389, adapter 63MiB.
  2. DPO v1 best: beta0.1 + NLL1.0 lr2e-5 grad_accum8, 607 pairs hellaswag 409 mmlu88 obqa78 gsm8k48 sciq27 arc25, 76 steps loss 13->7 ref logps precomputed. Screen 73.73% ID2.288 delta+1.92 vs SFT 71.81% W20 L15. v2/v3 regressed 72.5%.

magi-1.3-1.7B against LFM3-1.7B BF16: 3.8x less memory, 97% score, 3.9x per GiB

Capability density

Capability density: magi-1.3-1.7B 2.207, magi-1.2-2B 1.305, magi-1.1-2B 1.149, Qwen3.5-2B BF16 0.385, LFM3-1.7B BF16 0.613

rho = -log2(1-score)/GiB. Measured same way llama.cpp b11800 q4_0 KV -ub 128 8192 ctx per slot, memory at 32K from fit.log: 863 MiB = 700 model+100 ctx+63 compute. fit.log: CUDA0 RTX 5070 12226=10965+(863=700+100+63)+... Host 0.

Benchmarks

Protocol: 2313 tasks full + 250 screen, 12 suites, think 4096 greedy, q4_0 KV flash-attn ub128 num_ctx 8192 per slot, 6 lanes max efficiency 99 overall TPS @2.2GB (282 eval), 1 lane 6.3s/task 325 eval TPS 648 overall for stability, 7 lanes 2.5-2.9GB 65 overall 218 eval, 8 lanes deadlock 0 tasks @2698MiB 82% util (parallel scheduler circular wait). Base LFM3 BF16 3.3GB 3298MiB 16bpw, Q4_K_M 1.1GB, magi IQ2_XXS 871MiB 4.19bpw; Qwen3.5-2B BF16 4.72 GiB; magi-1.2-2B 826MB Ollama 789M GGUF 1.23GiB; magi-1.1-2B 1.36GiB.

BenchmarkQwen3.5-2B BF16
4.72 GiB
LFM3-1.7B BF16
3.22 GiB
magi-1.1-2B
1.36 GiB
magi-1.2-2B
1.23 GiB
magi-1.3-1.7B
0.84 GiB
Efficiency
Memory at 32K (GiB)4.723.221.361.230.84
Full2313 score71.66%~74.5% scr66.06%67.09%72.45%
Screen250 score~72% est74.54%66.1% est~67% est73.73%
Capability density rho0.3850.6131.1491.3052.207
Decode speed (tok/s) eval- ~300325333325
Mean think tokens3,018~1,6003,4092,4991,476
Hit 4K budget62.4%~35%80.7%48.2%13.5% 312/2313
Traces ending loop49.2%~20%77.4%38.8%~8% est
Coding
HumanEval+56.10~86% scr46.9550.6186.6
MBPP+55.64~68%44.6144.8668.2
LiveCodeBench11.33~25%1.338.0023.3
Math / Science
GSM8K82.67~91%78.0076.6790.7
ARC-Challenge94.67~90%93.3396.0088.7
MMLU74.50~60%64.0065.5059.0

vs last gen 2B: +5.36 pts vs magi-1.2-2B, +6.39 vs 1.1-2B, +~9 vs 1.0-2B, 0.39 GiB less, 1.69x density. vs base IFM LFM3: 74.54% BF16 ->72.45% =97% kept 3.8x less memory 3.6x density. vs Qwen base: 71.66%->72.45% +0.79 pts 5.6x less 5.7x density. LFM3 conv wins coding HE 86.6% vs 50.6% for 1.2-2B.

Extra graphs โ€” TPS + thinking efficiency

Thinking efficiency: magi-1.3-1.7B 1476 mean think, 13.5% budget hits, 6.3s/task vs magi-1.2-2B 2499/48.2%/8.2s vs magi-1.1-2B 3409/80.7%/12s

TPS: magi-1.3-1.7B 325 eval TPS 99 overall @2.2GB 6 lanes, magi-1.2-2B 333 eval 99 overall, magi-1.1-2B 325 eval, Qwen BF16 not measured, LFM3 BF16 ~300

Per-suite grouped bars for 3 models across 12 suites: HumanEval, MBPP, LCB, GSM8K, ARC, MMLU, OBQA, SciQ, HellaSwag, WinoGrande, TruthfulQA, ToolCall

Thinking tokens are mean over 2313 tasks: Qwen 3018, LFM3 ~1600, 1.1-2B 3409, 1.2-2B 2499, 1.3-1.7B 1476 โ€” 41% less than 1.2-2B, 57% less than 1.1-2B, due to SFT on short correct traces + DPO + LFM3 conv architecture (shortconv 7 L-cache reduces loop). Cap hits 13.5% vs 48.2% vs 80.7% โ€” DPO loop fix works. TPS: 1 lane 6.3s/task 325 eval, 6 lanes 2.2GB 99 overall 282 eval max efficiency (per 9GB ceiling), 7 lanes 2.5-2.9GB 65 overall 218 eval, 8 lanes deadlock 0 tasks 2698MiB โ€” Ollama parallel scheduler circular wait all slots prompt-eval. Must set OLLAMA_NUM_PARALLEL=6 before serve.

Recipe

Per-tensor types LFM3-1.7B 24 layers: 13 conv, 6 full_attn 2,6,10,14,18,22, 5 sliding w256.

Tensor group Types Share bpw vs 2B last gen 187 types vs 1.2-4B 249 types
FFN down/gate/up 24 layers IQ3_XXS x82 IQ4_XS x15 48% 3.3 Same 82 iq3_xxs as 2B +2 iq4_xs vs 2B -49 iq3_xxs vs 4B -4 iq4_xs
Token embd Q4_K x1 27% 4.5 vs q3_k in 2B (q3_k->q4_k) Same
Full-attn q,k,v,o+gate 6 layers Q4_K x31 12% 4.5 vs 37 q4_k +11 q5_k in 2B simpler no q5_k vs 32 q4_k +8 q5_k in 4B
ShortConv conv/in/out 13 layers Q8_0 x39 <1% 8.5 vs 36 q8_0 ssm in 2B +3 vs 48 q8_0 in 4B -9
Norms F32 <0.1% 32 Same Same

Quantizer llama-quantize --pure b4706 base IQ2_XXS override lfm3-magi.types.txt 167 (vs 187 for 2B, 249 for 4B). Imatrix imatrix-lfm3-own.gguf 3.5MB PPL2.3983 calib-4b.txt 22x8192.

No layers pruned: all 24 blk intact. Imatrix only quant importance.

Training

SFT LoRA r16 a32 w1,w2,w3,q_proj,k_proj,v_proj,o_proj,gate_proj,in_proj,out_proj 1696 commonsense-biased loss 0.668->0.389 adapter 63MiB chunked CE 128 QLoRA nf4 double quant bf16 checkpointing. DPO v1 beta0.1 NLL1.0 lr2e-5 grad_accum8 607 pairs 76 steps loss 13->7 ref logps precomputed. v2/v3 673/676 rows regressed.

How to run

llama.cpp: llama-server -m magi-1.3-1.7B-IQ2_XXS.gguf -c 32768 --jinja --flash-attn on --cache-type-k q4_0 --cache-type-v q4_0 -ub 128 -np 6

Ollama: $env:OLLAMA_NUM_PARALLEL='6'; $env:OLLAMA_MAX_QUEUE='12'; ollama serve; ollama create magi-1.3-1.7B -f Modelfile.lfm3-r36-dpo-v1-full; ollama run magi-1.3-1.7B

Limitations

LFM3 not Qwen: conv vs Gated DeltaNet cross-arch shows conv wins coding HE 86.6% vs 50.6% for 1.2-2B and GSM8K 90.7% vs 76.7%. Below BF16 screen 74.54%->72.45% 97% kept 2.1 pts gap hellaswag -10 winogrande -4.5, LCB 23.3% actually better than Qwen BF16? DPO helps. Still loops much less 13.5% vs 48%. Selected on screen250 v1 best confirmed full2313. Narrow eval one runtime.

Quant structure summary

Version Arch Layers Types GGUF BPW True Params Notable
magi-1.0-2B Qwen2.5 hybrid ~28 187 (82 iq3_xxs 37 q4_k 36 q8_0 13 iq4_xs 11 q5_k 7 iq2_s 1 q3_k) 789M ~3.4 1.88B q3_k token embd
magi-1.1-2B Qwen3.5-2B 24 187 same 895M ~3.5 1.88B +DPO loop
magi-1.2-2B Qwen3.5-2B 24 187 same breakdown 2B-fitted map 826M Ollama 789M ~3.4 1.88B 39 groups moved
magi-1.2-4B Qwen3.5-4B ~48 249 (131 iq3_xxs 48 q8_0 32 q4_k 19 iq4_xs 11 iq2_s 8 q5_k) 1.8G ~3.4 ~4.2B +62 types
magi-1.3-1.7B LFM3-1.7B conv 24 (13c 6f 5s) 167 (82 iq3_xxs 39 q8_0 31 q4_k 15 iq4_xs) 871M 913MB Ollama 4.19 1728801536 No iq2_s/q5_k/q3_k simpler higher BPW 863MiB fit

Related

GEHIRNS/magi-1.2-2B last gen 2B 67.09% @1.23 GiB rho1.305 187 types 826MB, magi-1.1-2B 66.06% @1.36 GiB rho1.149, magi-1.0-2B recipe alone, magi-1.2-4B 4B 249 types 1.8G M64x map, LiquidAI/LFM3-1.7B base 24 layers hybrid true 1728801536 128k vocab 262k max, Qwen/Qwen3.5-2B Qwen base 1.88B true, Paper Where the Bits Go DOI 10.5281/zenodo.23630199

License

Apache 2.0 inherited from LiquidAI/LFM3-1.7B. See base LICENSE.

Downloads last month
342
GGUF
Model size
2B params
Architecture
lfm3
Hardware compatibility
Log In to add your hardware

2-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for GEHIRNS/magi-1.3-1.7B

Finetuned
(87)
this model

Collection including GEHIRNS/magi-1.3-1.7B

Evaluation results