Clinical & Medical

Numeric reasoning, extraction, contract-holding and safety, measured on the file you ship.

How to read this table
  • Rank. Models clearly ahead, plus one; = is a tie. A row can rank better than a higher-scoring row above it, marked *, when its wider interval is cleared by fewer models.
  • Overall score. Percent of rubric points over 167 cases; a fired safety gate zeroes the case.
  • Judge. An LLM grades the open-ended answers; not used to rank.
  • Memory and speed. Measured on one desktop machine, not a phone; GGUF memory at phone settings; ONNX runs on the CPU.
  • iPhone 17 Pro+. Needs a 12 GB phone; other ranked rows fit 8 GB phones.
  • Fine-tune. Trained on undisclosed data, so overlap with these cases cannot be ruled out; not evidence of generalisation.
Model leaderboard · sorted by official order (overall score, compared by interval)
RankModel
1=89.886.2–93.07.05 GB92.87.05 GB7.25 GiB76 tok/s
1=86.282.1–89.92.68 GB90.62.68 GB3.38 GiB85 tok/s
1=86.181.5–90.31.94 GB90.31.94 GB1.99 GiB183 tok/s
1=86.081.8–89.83.54 GB89.43.54 GB3.83 GiB111 tok/s
1=82.878.5–86.81.32 GB86.71.32 GB1.28 GiB98 tok/s
1=82.377.9–86.52.25 GB78.92.25 GB2.51 GiB141 tok/s
2=81.378.0–84.32.72 GB93.32.72 GB3.21 GiB138 tok/s
2=81.176.4–85.52.61 GB87.22.61 GB2.90 GiB118 tok/s
2=80.575.7–85.03.58 GB84.43.58 GB4.10 GiB129 tok/s
2=80.575.8–84.91.94 GB81.11.94 GB2.16 GiB93 tok/s
4=77.372.6–81.72.60 GB81.72.60 GB2.96 GiB100 tok/s
4=76.871.9–81.52.26 GB79.42.26 GB1.75 GiB168 tok/s
5=76.071.1–80.81.77 GB71.71.77 GB2.30 GiBCPU26 tok/sCPU
5=74.169.2–78.82.68 GB71.72.68 GB2.76 GiB132 tok/s
673.468.2–78.40.81 GB73.30.81 GB0.93 GiB435 tok/s
970.965.4–76.22.10 GB85.02.10 GB3.09 GiBCPU68 tok/sCPU
11=70.766.0–75.22.78 GB80.02.78 GB3.00 GiB118 tok/s
11=68.363.5–73.21.25 GB70.61.25 GB1.23 GiB206 tok/s
11=68.162.7–73.41.47 GB71.11.47 GB1.93 GiB192 tok/s
11=67.662.3–72.81.45 GB71.11.45 GB1.54 GiB222 tok/s
1464.959.1–70.61.82 GB77.81.82 GB2.05 GiB175 tok/s
1563.057.4–68.41.21 GB59.41.21 GB1.53 GiB171 tok/s
1662.556.8–68.11.05 GB76.71.05 GB1.23 GiB229 tok/s
18=60.054.6–65.21.50 GB68.31.50 GB1.64 GiB334 tok/s
18=59.353.7–64.90.66 GB66.10.66 GB0.79 GiB424 tok/s
18=59.353.5–65.00.99 GB47.80.99 GB1.23 GiBCPU89 tok/sCPU
2156.251.1–61.32.19 GB57.82.19 GB2.54 GiB157 tok/s
2450.544.5–56.31.45 GB64.41.45 GB1.55 GiB207 tok/s
2844.839.2–50.50.73 GB33.90.73 GB0.81 GiB355 tok/s

What does compression cost?

Score by bits per weight: MLX solid, GGUF dashed, ONNX dotted; ring = build to ship, red = cliff, amber = isolated drop.

Gemma 4 E4B ITcliff at MLX 4-bit, −6.0
306510016.07 bpw2.30 bpwGemma 4 E4B IT · 8-bit MLX · 8.95 bpw · 88.2% measured · 8.91 GB · reference rungGemma 4 E4B IT · 6-bit MLX · 7.07 bpw · 89.8% measured · 7.05 GB · build to shipGemma 4 E4B IT · 5-bit MLX · 6.13 bpw · 87.9% measured · 6.11 GBGemma 4 E4B IT · 4-bit MLX · 5.18 bpw · 82.2% measured · 5.18 GB · below the family's MLX cliffGemma 4 E4B IT · INT4 ONNX · 5.57 bpw · 86.0% measured · 5.94 GB
LFM2.5 2.6B
306510016.07 bpw2.30 bpwLFM2.5 2.6B · Q8_0 GGUF · 8.50 bpw · 89.3% measured · 2.87 GB · reference rung · 4 of 167 answers cut off at the budgetLFM2.5 2.6B · Q6_K GGUF · 6.57 bpw · 88.1% measured · 2.22 GB · 8 of 167 answers cut off at the budgetLFM2.5 2.6B · Q5_K_M GGUF · 5.73 bpw · 86.1% measured · 1.94 GB · build to ship · 6 of 167 answers cut off at the budgetLFM2.5 2.6B · Q4_K_M GGUF · 4.94 bpw · 87.1% measured · 1.67 GB · 9 of 167 answers cut off at the budgetLFM2.5 2.6B · 8-bit MLX · 8.50 bpw · 87.9% measured · 2.88 GB · 5 of 167 answers cut off at the budgetLFM2.5 2.6B · 6-bit MLX · 6.50 bpw · 87.5% measured · 2.21 GB · 8 of 167 answers cut off at the budgetLFM2.5 2.6B · 5-bit MLX · 5.60 bpw · 87.1% measured · 1.91 GB · 5 of 167 answers cut off at the budgetLFM2.5 2.6B · 4-bit MLX · 4.70 bpw · 85.1% measured · 1.60 GB · 6 of 167 answers cut off at the budgetLFM2.5 2.6B · INT4 ONNX · 5.40 bpw · 85.0% measured · 1.87 GB · 7 of 167 answers cut off at the budget
Nanbeige4.2 3B
306510016.07 bpw2.30 bpwNanbeige4.2 3B · BF16 GGUF · 16.00 bpw · 86.1% measured · 8.34 GB · reference rungNanbeige4.2 3B · Q8_0 GGUF · 8.50 bpw · 86.4% measured · 4.43 GBNanbeige4.2 3B · Q6_K GGUF · 6.89 bpw · 86.3% measured · 3.60 GBNanbeige4.2 3B · Q5_K_M GGUF · 5.89 bpw · 87.1% measured · 3.08 GBNanbeige4.2 3B · Q4_K_M GGUF · 5.14 bpw · 86.2% measured · 2.68 GB · build to ship
NVIDIA Nemotron 3 Nano 4B
306510016.07 bpw2.30 bpwNVIDIA Nemotron 3 Nano 4B · Q8_0 GGUF · 8.51 bpw · 83.9% measured · 4.23 GB · reference rungNVIDIA Nemotron 3 Nano 4B · Q6_K GGUF · 7.90 bpw · 84.2% measured · 3.93 GBNVIDIA Nemotron 3 Nano 4B · Q4_K_M GGUF · 5.70 bpw · 83.1% measured · 2.84 GBNVIDIA Nemotron 3 Nano 4B · 8-bit MLX · 8.50 bpw · 85.5% measured · 4.24 GBNVIDIA Nemotron 3 Nano 4B · 6-bit MLX · 6.50 bpw · 85.1% measured · 3.25 GBNVIDIA Nemotron 3 Nano 4B · 4-bit MLX · 4.50 bpw · 82.3% measured · 2.25 GB · build to shipNVIDIA Nemotron 3 Nano 4B · INT4 ONNX · 5.14 bpw · 82.1% measured · 2.57 GB
Spark X2.6 4B
306510016.07 bpw2.30 bpwSpark X2.6 4B · BF16 GGUF · 16.00 bpw · 85.1% measured · 8.23 GB · reference rungSpark X2.6 4B · Q8_0 GGUF · 8.50 bpw · 86.1% measured · 4.38 GBSpark X2.6 4B · Q4_K_M GGUF · 6.36 bpw · 86.0% measured · 3.54 GB · build to ship
Ternary Bonsai 2 4B
306510016.07 bpw2.30 bpwTernary Bonsai 2 4B · 2-bit MLX · 3.30 bpw · 82.7% measured · 1.88 GB · reference rungTernary Bonsai 2 4B · Q2_0 GGUF · 2.76 bpw · 82.8% measured · 1.58 GBTernary Bonsai 2 4B · PQ2_0 GGUF · 2.65 bpw · 82.8% measured · 1.51 GBTernary Bonsai 2 4B · PTQ1_0 GGUF · 2.30 bpw · 82.8% measured · 1.32 GB · build to ship
Gemma 4 E2B IT
306510016.07 bpw2.30 bpwGemma 4 E2B IT · 8-bit MLX · 9.19 bpw · 80.2% measured · 5.90 GB · reference rungGemma 4 E2B IT · 6-bit MLX · 7.38 bpw · 79.4% measured · 4.74 GBGemma 4 E2B IT · 5-bit MLX · 6.47 bpw · 76.9% measured · 4.16 GBGemma 4 E2B IT · 4-bit MLX · 5.56 bpw · 80.5% measured · 3.58 GB · build to shipGemma 4 E2B IT · INT4 ONNX · 5.54 bpw · 78.8% measured · 3.95 GB
Granite 4.0 H Micro
306510016.07 bpw2.30 bpwGranite 4.0 H Micro · BF16 GGUF · 16.01 bpw · 83.0% measured · 6.39 GB · reference rungGranite 4.0 H Micro · Q8_0 GGUF · 8.51 bpw · 83.6% measured · 3.40 GBGranite 4.0 H Micro · Q6_K GGUF · 6.57 bpw · 82.0% measured · 2.63 GBGranite 4.0 H Micro · Q5_K_M GGUF · 5.69 bpw · 80.6% measured · 2.27 GBGranite 4.0 H Micro · Q4_K_M GGUF · 4.86 bpw · 80.5% measured · 1.94 GB · build to shipGranite 4.0 H Micro · INT4 ONNX · 5.40 bpw · 78.4% measured · 2.16 GB
Granite 4.2 3Bcliff at MLX 4-bit, −8.3
306510016.07 bpw2.30 bpwGranite 4.2 3B · BF16 GGUF · 16.00 bpw · 82.4% measured · 7.32 GB · reference rungGranite 4.2 3B · Q8_0 GGUF · 8.50 bpw · 82.6% measured · 3.89 GBGranite 4.2 3B · Q6_K GGUF · 6.56 bpw · 81.0% measured · 3.01 GBGranite 4.2 3B · Q5_K_M GGUF · 5.71 bpw · 81.1% measured · 2.61 GB · build to shipGranite 4.2 3B · Q4_K_M GGUF · 4.90 bpw · 79.2% measured · 2.24 GBGranite 4.2 3B · BF16 MLX · 16.00 bpw · 82.6% measured · 7.33 GBGranite 4.2 3B · 8-bit MLX · 8.50 bpw · 81.7% measured · 3.90 GBGranite 4.2 3B · 4-bit MLX · 4.50 bpw · 74.1% measured · 2.07 GB · below the family's MLX cliff
MedPsy 4B
306510016.07 bpw2.30 bpwMedPsy 4B · BF16 GGUF · 16.00 bpw · 78.3% measured · 8.83 GB · reference rungMedPsy 4B · Q8_0 GGUF · 8.50 bpw · 80.0% measured · 4.69 GBMedPsy 4B · Q5_K_M-imat GGUF · 5.71 bpw · 79.8% measured · 3.16 GBMedPsy 4B · Q4_K_M-imat GGUF · 4.91 bpw · 81.3% measured · 2.72 GB · build to shipMedPsy 4B · INT4 mixed ONNX · 6.28 bpw · 79.8% measured · 3.31 GB
LFM2.5 VL 3B
306510016.07 bpw2.30 bpwLFM2.5 VL 3B · BF16 MLX · 16.00 bpw · 76.5% measured · 6.27 GB · reference rungLFM2.5 VL 3B · 8-bit MLX · 9.52 bpw · 75.4% measured · 3.74 GBLFM2.5 VL 3B · 6-bit MLX · 7.80 bpw · 76.8% measured · 3.06 GBLFM2.5 VL 3B · 5-bit MLX · 7.02 bpw · 76.7% measured · 2.76 GBLFM2.5 VL 3B · 4-bit MLX · 6.07 bpw · 77.8% measured · 2.39 GBLFM2.5 VL 3B · INT4 + FP32 ONNX · 9.96 bpw · 75.4% measured · 3.94 GBLFM2.5 VL 3B · Q8_0 GGUF · 8.84 bpw · 76.1% measured · 3.46 GBLFM2.5 VL 3B · Q6_K GGUF · 7.16 bpw · 76.7% measured · 2.80 GBLFM2.5 VL 3B · Q5_K_M GGUF · 6.44 bpw · 78.3% measured · 2.52 GBLFM2.5 VL 3B · Q4_K_M GGUF · 5.76 bpw · 76.8% measured · 2.26 GB · build to ship
Spark X2.5 4B
306510016.07 bpw2.30 bpwSpark X2.5 4B · BF16 GGUF · 16.00 bpw · 79.9% measured · 8.23 GB · reference rungSpark X2.5 4B · Q8_0 GGUF · 8.50 bpw · 80.9% measured · 4.38 GBSpark X2.5 4B · Q4_K_M GGUF · 5.05 bpw · 77.3% measured · 2.60 GB · build to ship
MiniCPM5 2Bisolated drop at Q4_K_M, 2 rungs below
306510016.07 bpw2.30 bpwMiniCPM5 2B · F16 GGUF · 16.00 bpw · 74.3% measured · 5.04 GB · reference rungMiniCPM5 2B · Q8_0 GGUF · 8.50 bpw · 74.1% measured · 2.68 GB · build to shipMiniCPM5 2B · Q4_K_M GGUF · 4.95 bpw · 68.6% measured · 1.56 GB · isolated drop (not a cliff: a narrower rung of this format holds)MiniCPM5 2B · INT4-F16 ONNX · 5.72 bpw · 68.2% measured · 1.84 GBMiniCPM5 2B · 4-bit MLX · 4.50 bpw · 64.0% measured · 1.43 GB · isolated drop (not a cliff: a narrower rung of this format holds)
Qwen3.5 2Bcliff at MLX 4-bit, −8.4
306510016.07 bpw2.30 bpwQwen3.5 2B · BF16 GGUF · 16.02 bpw · 78.3% measured · 4.44 GB · reference rungQwen3.5 2B · Q8_0 GGUF · 9.65 bpw · 79.3% measured · 2.68 GBQwen3.5 2B · Q6_K GGUF · 8.07 bpw · 77.3% measured · 2.24 GBQwen3.5 2B · Q5_K_M GGUF · 7.56 bpw · 77.4% measured · 2.10 GBQwen3.5 2B · Q4_K_M GGUF · 7.01 bpw · 77.6% measured · 1.95 GBQwen3.5 2B · 8-bit MLX · 9.62 bpw · 79.1% measured · 2.69 GBQwen3.5 2B · 6-bit MLX · 7.92 bpw · 79.8% measured · 2.22 GBQwen3.5 2B · 5-bit MLX · 7.07 bpw · 77.6% measured · 1.98 GBQwen3.5 2B · 4-bit MLX · 6.22 bpw · 69.9% measured · 1.75 GB · below the family's MLX cliffQwen3.5 2B · INT4 ONNX · 6.32 bpw · 76.0% measured · 1.77 GB · build to ship
LFM2.5 1.2B Thinkingcliff at MLX 4-bit, −6.0
306510016.07 bpw2.30 bpwLFM2.5 1.2B Thinking · BF16 MLX · 16.00 bpw · 74.4% measured · 2.35 GB · reference rung · 8 of 167 answers cut off at the budgetLFM2.5 1.2B Thinking · 8-bit MLX · 8.50 bpw · 75.9% measured · 1.25 GB · 7 of 167 answers cut off at the budgetLFM2.5 1.2B Thinking · 6-bit MLX · 6.50 bpw · 74.7% measured · 0.96 GB · 5 of 167 answers cut off at the budgetLFM2.5 1.2B Thinking · 5-bit MLX · 5.50 bpw · 73.4% measured · 0.81 GB · build to ship · 8 of 167 answers cut off at the budgetLFM2.5 1.2B Thinking · 4-bit MLX · 4.50 bpw · 68.4% measured · 0.66 GB · 12 of 167 answers cut off at the budget · below the family's MLX cliffLFM2.5 1.2B Thinking · INT4 ONNX · 5.59 bpw · 70.1% measured · 0.85 GB · 9 of 167 answers cut off at the budget
LFM2 2.6Bisolated drop at Q4_K_M, 2 rungs below
306510016.07 bpw2.30 bpwLFM2 2.6B · F16 GGUF · 16.00 bpw · 75.2% measured · 5.14 GB · reference rungLFM2 2.6B · Q8_0 GGUF · 8.50 bpw · 74.2% measured · 2.73 GBLFM2 2.6B · Q6_K GGUF · 6.57 bpw · 73.6% measured · 2.11 GBLFM2 2.6B · Q5_K_M GGUF · 5.69 bpw · 70.7% measured · 1.83 GBLFM2 2.6B · Q4_K_M GGUF · 4.86 bpw · 68.8% measured · 1.56 GB · isolated drop (not a cliff: a narrower rung of this format holds)LFM2 2.6B · 8-bit MLX · 8.50 bpw · 72.4% measured · 2.74 GBLFM2 2.6B · 4-bit MLX · 4.50 bpw · 67.3% measured · 1.45 GB · isolated drop (not a cliff: a narrower rung of this format holds)LFM2 2.6B · INT4 + FP32 ONNX · 6.41 bpw · 70.9% measured · 2.10 GB · build to ship
LFM2 2.6B Expisolated drop at 5-bit, 2 rungs below
306510016.07 bpw2.30 bpwLFM2 2.6B Exp · F16 GGUF · 16.00 bpw · 69.6% measured · 5.41 GB · reference rungLFM2 2.6B Exp · Q8_0 GGUF · 8.50 bpw · 69.4% measured · 2.88 GBLFM2 2.6B Exp · Q6_K GGUF · 6.57 bpw · 70.0% measured · 2.22 GBLFM2 2.6B Exp · Q5_K_M GGUF · 5.68 bpw · 70.6% measured · 1.92 GBLFM2 2.6B Exp · Q4_K_M GGUF · 4.84 bpw · 63.8% measured · 1.64 GB · isolated drop (not a cliff: a narrower rung of this format holds)LFM2 2.6B Exp · BF16 MLX · 16.00 bpw · 68.9% measured · 5.14 GBLFM2 2.6B Exp · 8-bit MLX · 8.50 bpw · 69.5% measured · 2.74 GBLFM2 2.6B Exp · 6-bit MLX · 6.50 bpw · 69.0% measured · 2.09 GBLFM2 2.6B Exp · 5-bit MLX · 5.50 bpw · 62.9% measured · 1.77 GB · isolated drop (not a cliff: a narrower rung of this format holds)LFM2 2.6B Exp · 4-bit MLX · 4.50 bpw · 67.6% measured · 1.45 GB · build to shipLFM2 2.6B Exp · INT4 ONNX · 5.34 bpw · 64.9% measured · 1.75 GB
LFM3 VL 2B
306510016.07 bpw2.30 bpwLFM3 VL 2B · BF16 MLX · 16.00 bpw · 70.4% measured · 3.85 GB · reference rungLFM3 VL 2B · 8-bit MLX · 9.24 bpw · 69.8% measured · 2.23 GBLFM3 VL 2B · 6-bit MLX · 7.44 bpw · 67.3% measured · 1.80 GBLFM3 VL 2B · 4-bit MLX · 6.63 bpw · 68.2% measured · 1.61 GBLFM3 VL 2B · 5-bit MLX · 6.54 bpw · 71.0% measured · 1.59 GBLFM3 VL 2B · BF16 GGUF · 15.27 bpw · 70.1% measured · 3.67 GBLFM3 VL 2B · Q8_0 GGUF · 8.51 bpw · 70.0% measured · 2.05 GBLFM3 VL 2B · Q6_K GGUF · 6.77 bpw · 71.3% measured · 1.63 GBLFM3 VL 2B · Q5_K_M GGUF · 6.06 bpw · 67.3% measured · 1.46 GBLFM3 VL 2B · Q4_K_M GGUF · 5.40 bpw · 66.3% measured · 1.30 GBLFM3 VL 2B · QAD-Q4_0 GGUF · 5.19 bpw · 68.3% measured · 1.25 GB · build to ship
MedPsy 1.7Bcliff at GGUF Q4_K_M-imat, −12.6
306510016.07 bpw2.30 bpwMedPsy 1.7B · BF16 GGUF · 16.00 bpw · 69.2% measured · 4.07 GB · reference rungMedPsy 1.7B · Q8_0 GGUF · 8.50 bpw · 69.4% measured · 2.17 GBMedPsy 1.7B · Q5_K_M-imat GGUF · 5.77 bpw · 68.1% measured · 1.47 GB · build to shipMedPsy 1.7B · Q4_K_M-imat GGUF · 5.03 bpw · 56.6% measured · 1.28 GB · below the family's GGUF cliffMedPsy 1.7B · INT4 mixed ONNX · 6.58 bpw · 67.9% measured · 1.45 GB
Ministral 3 3B Instruct 2512
306510016.07 bpw2.30 bpwMinistral 3 3B Instruct 2512 · BF16 GGUF · 16.00 bpw · 71.4% measured · 7.71 GB · reference rungMinistral 3 3B Instruct 2512 · Q8_0 GGUF · 9.32 bpw · 72.1% measured · 4.49 GBMinistral 3 3B Instruct 2512 · Q5_K_M GGUF · 6.87 bpw · 70.0% measured · 3.32 GBMinistral 3 3B Instruct 2512 · Q4_K_M GGUF · 6.19 bpw · 69.3% measured · 2.99 GBMinistral 3 3B Instruct 2512 · BF16 MLX · 16.00 bpw · 71.2% measured · 7.73 GBMinistral 3 3B Instruct 2512 · 8-bit MLX · 9.29 bpw · 71.0% measured · 4.50 GBMinistral 3 3B Instruct 2512 · 6-bit MLX · 7.50 bpw · 68.2% measured · 3.64 GBMinistral 3 3B Instruct 2512 · 5-bit MLX · 6.60 bpw · 69.8% measured · 3.21 GBMinistral 3 3B Instruct 2512 · 4-bit MLX · 5.71 bpw · 70.7% measured · 2.78 GB · build to shipMinistral 3 3B Instruct 2512 · INT4 ONNX · 5.56 bpw · 68.1% measured · 2.82 GB
Llama 3.2 3B Instruct
306510016.07 bpw2.30 bpwLlama 3.2 3B Instruct · BF16 MLX · 16.00 bpw · 65.8% measured · 6.43 GB · reference rungLlama 3.2 3B Instruct · 8-bit MLX · 8.50 bpw · 63.6% measured · 3.42 GBLlama 3.2 3B Instruct · 4-bit MLX · 4.50 bpw · 64.9% measured · 1.82 GB · build to shipLlama 3.2 3B Instruct · Q8_0 GGUF · 8.50 bpw · 64.9% measured · 3.42 GBLlama 3.2 3B Instruct · Q6_K GGUF · 6.56 bpw · 62.4% measured · 2.64 GBLlama 3.2 3B Instruct · Q5_K_M GGUF · 5.76 bpw · 63.9% measured · 2.32 GBLlama 3.2 3B Instruct · Q4_K_M GGUF · 5.01 bpw · 66.6% measured · 2.02 GBLlama 3.2 3B Instruct · INT4 + FP32 ONNX · 8.31 bpw · 68.4% measured · 3.42 GB
Ternary Bonsai 2 2B
306510016.07 bpw2.30 bpwTernary Bonsai 2 2B · 2-bit MLX · 4.34 bpw · 63.0% measured · 1.21 GB · reference rung · build to shipTernary Bonsai 2 2B · Q2_0 GGUF · 3.25 bpw · 61.0% measured · 0.91 GBTernary Bonsai 2 2B · PQ2_0 GGUF · 3.15 bpw · 61.0% measured · 0.88 GBTernary Bonsai 2 2B · PTQ1_0 GGUF · 2.83 bpw · 61.0% measured · 0.79 GB
LFM3 1.7Bcliff at MLX 4-bit, −9.8
306510016.07 bpw2.30 bpwLFM3 1.7B · BF16 GGUF · 16.00 bpw · 65.8% measured · 3.47 GB · reference rungLFM3 1.7B · Q8_0 GGUF · 8.51 bpw · 65.6% measured · 1.85 GBLFM3 1.7B · Q6_K GGUF · 6.57 bpw · 65.9% measured · 1.43 GBLFM3 1.7B · Q5_K_M GGUF · 5.79 bpw · 66.2% measured · 1.26 GBLFM3 1.7B · Q4_K_M GGUF · 5.05 bpw · 63.8% measured · 1.10 GBLFM3 1.7B · QAD-Q4_0 GGUF · 4.82 bpw · 62.5% measured · 1.05 GB · build to shipLFM3 1.7B · BF16 MLX · 16.00 bpw · 65.8% measured · 3.48 GB · 1 of 167 answers cut off at the budgetLFM3 1.7B · 8-bit MLX · 8.50 bpw · 65.7% measured · 1.86 GB · 2 of 167 answers cut off at the budgetLFM3 1.7B · 6-bit MLX · 6.50 bpw · 64.5% measured · 1.42 GB · 1 of 167 answers cut off at the budgetLFM3 1.7B · 5-bit MLX · 5.50 bpw · 62.4% measured · 1.21 GB · 1 of 167 answers cut off at the budgetLFM3 1.7B · 4-bit MLX · 4.50 bpw · 56.0% measured · 0.99 GB · below the family's MLX cliff
LFM2.5 1.2B Instructdrop −7.2, not significant
306510016.07 bpw2.30 bpwLFM2.5 1.2B Instruct · BF16 GGUF · 16.00 bpw · 60.0% measured · 2.34 GB · reference rungLFM2.5 1.2B Instruct · Q8_0 GGUF · 8.50 bpw · 60.6% measured · 1.25 GBLFM2.5 1.2B Instruct · Q6_K GGUF · 6.57 bpw · 60.3% measured · 0.96 GBLFM2.5 1.2B Instruct · Q5_K_M GGUF · 5.75 bpw · 59.5% measured · 0.84 GBLFM2.5 1.2B Instruct · Q4_K_M GGUF · 4.98 bpw · 60.0% measured · 0.73 GBLFM2.5 1.2B Instruct · BF16 MLX · 16.00 bpw · 59.9% measured · 2.35 GBLFM2.5 1.2B Instruct · 8-bit MLX · 8.50 bpw · 59.2% measured · 1.25 GBLFM2.5 1.2B Instruct · 6-bit MLX · 6.50 bpw · 56.4% measured · 0.96 GBLFM2.5 1.2B Instruct · 5-bit MLX · 5.50 bpw · 52.8% measured · 0.81 GBLFM2.5 1.2B Instruct · 4-bit MLX · 4.50 bpw · 59.3% measured · 0.66 GB · build to shipLFM2.5 1.2B Instruct · INT4 ONNX · 5.59 bpw · 60.1% measured · 0.85 GB
LFM2.5 VL 1.6B
306510016.07 bpw2.30 bpwLFM2.5 VL 1.6B · BF16 MLX · 16.00 bpw · 62.9% measured · 3.20 GB · reference rungLFM2.5 VL 1.6B · 8-bit MLX · 10.44 bpw · 64.0% measured · 2.09 GBLFM2.5 VL 1.6B · 6-bit MLX · 8.96 bpw · 65.3% measured · 1.79 GBLFM2.5 VL 1.6B · 5-bit MLX · 8.21 bpw · 64.4% measured · 1.64 GBLFM2.5 VL 1.6B · 4-bit MLX · 7.47 bpw · 60.0% measured · 1.50 GB · build to shipLFM2.5 VL 1.6B · BF16 GGUF · 14.65 bpw · 63.9% measured · 2.93 GBLFM2.5 VL 1.6B · Q8_0 GGUF · 9.15 bpw · 63.9% measured · 1.83 GBLFM2.5 VL 1.6B · Q6_K GGUF · 7.73 bpw · 64.8% measured · 1.55 GB
MiniCPM V 4.7 1B
306510016.07 bpw2.30 bpwMiniCPM V 4.7 1B · F16 GGUF · 16.07 bpw · 60.5% measured · 3.13 GB · reference rungMiniCPM V 4.7 1B · Q8_0 GGUF · 11.22 bpw · 61.3% measured · 2.19 GBMiniCPM V 4.7 1B · Q6_K GGUF · 9.97 bpw · 60.4% measured · 1.95 GBMiniCPM V 4.7 1B · Q5_K_M GGUF · 9.53 bpw · 59.3% measured · 1.86 GBMiniCPM V 4.7 1B · Q4_K_M GGUF · 9.12 bpw · 57.3% measured · 1.78 GBMiniCPM V 4.7 1B · FP16, text only ONNX · 16.00 bpw · 61.0% measured · 2.07 GBMiniCPM V 4.7 1B · INT4-F16, text only ONNX · 7.41 bpw · 59.3% measured · 0.99 GB · build to shipMiniCPM V 4.7 1B · BF16 MLX · 16.00 bpw · 60.3% measured · 3.13 GBMiniCPM V 4.7 1B · 8-bit MLX · 14.51 bpw · 59.0% measured · 2.84 GBMiniCPM V 4.7 1B · 6-bit MLX · 14.12 bpw · 59.9% measured · 2.76 GBMiniCPM V 4.7 1B · 5-bit MLX · 13.92 bpw · 59.9% measured · 2.72 GBMiniCPM V 4.7 1B · 4-bit MLX · 13.72 bpw · 57.5% measured · 2.69 GB
North Micro Vision Instruct
306510016.07 bpw2.30 bpwNorth Micro Vision Instruct · BF16 MLX · 16.00 bpw · 59.5% measured · 4.99 GB · reference rungNorth Micro Vision Instruct · 8-bit MLX · 10.13 bpw · 58.9% measured · 3.16 GBNorth Micro Vision Instruct · 6-bit MLX · 8.56 bpw · 59.3% measured · 2.68 GBNorth Micro Vision Instruct · 5-bit MLX · 7.78 bpw · 58.1% measured · 2.44 GBNorth Micro Vision Instruct · 4-bit MLX · 6.99 bpw · 56.2% measured · 2.19 GB · build to ship
LFM2 2.6B Longevity
306510016.07 bpw2.30 bpwLFM2 2.6B Longevity · F16 GGUF · 16.00 bpw · 48.1% measured · 5.14 GB · reference rungLFM2 2.6B Longevity · Q8_0 GGUF · 8.50 bpw · 49.9% measured · 2.73 GBLFM2 2.6B Longevity · Q6_K GGUF · 6.57 bpw · 50.8% measured · 2.11 GBLFM2 2.6B Longevity · Q5_K_M GGUF · 5.69 bpw · 48.8% measured · 1.83 GBLFM2 2.6B Longevity · Q4_K_M GGUF · 4.86 bpw · 51.7% measured · 1.56 GBLFM2 2.6B Longevity · 8-bit MLX · 8.50 bpw · 48.9% measured · 2.73 GBLFM2 2.6B Longevity · 4-bit MLX · 4.50 bpw · 50.5% measured · 1.45 GB · build to shipLFM2 2.6B Longevity · INT4 mixed ONNX · 5.84 bpw · 47.4% measured · 1.92 GB
LFM2 1.2B Longevitycliff at MLX 4-bit, −5.2
306510016.07 bpw2.30 bpwLFM2 1.2B Longevity · F16 GGUF · 16.00 bpw · 48.1% measured · 2.34 GB · reference rungLFM2 1.2B Longevity · Q8_0 GGUF · 8.50 bpw · 46.1% measured · 1.25 GBLFM2 1.2B Longevity · Q6_K GGUF · 6.57 bpw · 46.1% measured · 0.96 GBLFM2 1.2B Longevity · Q5_K_M GGUF · 5.75 bpw · 47.6% measured · 0.84 GBLFM2 1.2B Longevity · Q4_K_M GGUF · 4.98 bpw · 44.8% measured · 0.73 GB · build to shipLFM2 1.2B Longevity · 8-bit MLX · 8.50 bpw · 45.8% measured · 1.25 GBLFM2 1.2B Longevity · 4-bit MLX · 4.50 bpw · 42.9% measured · 0.66 GB · below the family's MLX cliffLFM2 1.2B Longevity · INT4 mixed ONNX · 6.08 bpw · 47.5% measured · 0.93 GB

What does a phone's budget buy?

Score by download size: ringed = the build each family ships, grey = its other builds, hollow = over 4 GB; hover to trace a family, drag to zoom.

4050607080901001 GB2 GB4 GB8 GB4 GB phone budgetNanbeige4.2 3B · Q4_K_MLFM2.5 2.6B · Q5_K_MGemma 4 E4B IT · 6-bitTernary Bonsai 2 4B · PTQ1_0LFM2.5 1.2B Thinking · 5-bitLFM2.5 1.2B Instruct · 4-bitdownload size (log scale), vision tower included