QuantDiff

Best Sage-A1.5 quant for your GPU

4 quantized files of InternScience/Sage-A1.5 (35.51B): accuracy, tool calls and speed (no KLD: reference too large).

InternScience/Sage-A1.5 (Apache 2.0). Released Oct 21, 2026 by Shanghai AI Laboratory (InternScience). Agentic mixture-of-experts model built on Qwen3.6 35B-A3B: about 36B total, 3B active parameters per token; 40 layers (10 full attention + 30 linear attention), 256 experts with 8 selected, so the KV cache is small. Multimodal (image input via a separate projector), but we tested text only (projector not loaded). Tested with llama.cpp build b11869; tool calls and accuracy tested with thinking disabled. GGUF quants: the official InternScience Q4_K_M (made without an imatrix), mradermacher i1 imatrix quants, and na0x2c6's imatrix IQ4_XS-exps mix. None of the GGUF files we found on Oct 25 fit a 12 GB card (the smallest, mradermacher i1-Q2_K, is 13.2 GB).

Accuracy and speed only, no KLD: reference too large. The BF16 reference (about 72 GB) does not fit in our 32 GB test Mac (minipro), so there is no KLD for this model; we measured accuracy, tool calls and speed only. We never estimate KLD; picks for this model are ranked by measured accuracy (GSM8K + MMLU + ARC-C, 500 items, 95% CI) with tool calls as tie-break.

Charts

Every dot and bar is a measured file (GGUF). Hover or tap one for its numbers; each chart has a table under "Show the numbers". Download a PNG to share.

Sage-A1.5: accuracy vs file size

If a dot's bar reaches the shaded band, that file scores the same as the best one within the margin of error on our accuracy questions (GSM8K + MMLU + ARC-C); the smallest such file here is i1-Q2_K at 13.2 GB.

Sage-A1.5: accuracy vs file sizeScatter of 4 GGUF files of Sage-A1.5: file size in GB across, accuracy in percent up, with 95% interval bars. Shaded band: the best file's 95% interval (87.3 to 92.5%). Data table follows.82.585.087.590.092.595.012141618202224File size (GB)Accuracy (% of 500 questions)i1-Q2_K by mradermacher · 13.2 GB · accuracy 86.8% (83.6–89.5) · tied with the bestQ4_K_M by InternScience · 21.7 GB · accuracy 89.8% (86.8–92.2) · tied with the besti1-IQ3_M by mradermacher · 15.8 GB · accuracy 88.0% (84.9–90.6) · tied with the best · our 16 GB pickIQ4_XS-exps by na0x2c6 · 19.3 GB · accuracy 90.2% (87.3–92.5) · tied with the best · our 24 GB picki1-IQ3_MIQ4_XS-expsi1-Q2_KQ4_K_M
  • InternScience
  • mradermacher
  • na0x2c6
  • best file's 95% range
  • clearly below the best
  • our pick for 8 / 12 / 16 GB GPUs

Up: in %, bars = · Across: file size in GB · = overlaps the band

Show the numbers (table)
QuantGBAccuracy %95% intervalCorrectTied with bestPick
i1-Q2_K13.286.883.6–89.5434/500yes
i1-IQ3_M15.888.084.9–90.6440/500yes16 GB pick
IQ4_XS-exps19.390.287.3–92.5451/500yes24 GB pick
Q4_K_M21.789.886.8–92.2449/500yes

Sage-A1.5: tool calls vs file size

If a dot's bar reaches the shaded band, that file scores the same as the best one within the margin of error on our tool-call tests; the smallest such file here is i1-Q2_K at 13.2 GB.

Sage-A1.5: tool calls vs file sizeScatter of 4 GGUF files of Sage-A1.5: file size in GB across, tool calls in percent up, with 95% interval bars. Shaded band: the best file's 95% interval (81.6 to 91.0%). Data table follows.758085909512141618202224File size (GB)Tool calls (% of 200 cases)i1-Q2_K by mradermacher · 13.2 GB · tool calls 82.0% (76.1–86.7) · tied with the bestQ4_K_M by InternScience · 21.7 GB · tool calls 85.0% (79.4–89.3) · tied with the besti1-IQ3_M by mradermacher · 15.8 GB · tool calls 84.0% (78.3–88.4) · tied with the best · our 16 GB pickIQ4_XS-exps by na0x2c6 · 19.3 GB · tool calls 87.0% (81.6–91.0) · tied with the best · our 24 GB picki1-IQ3_MIQ4_XS-expsi1-Q2_KQ4_K_M
  • InternScience
  • mradermacher
  • na0x2c6
  • best file's 95% range
  • clearly below the best
  • our pick for 8 / 12 / 16 GB GPUs

Up: in %, bars = · Across: file size in GB · = overlaps the band

Show the numbers (table)
QuantGBTool calls %95% intervalCorrectTied with bestPick
i1-Q2_K13.282.076.1–86.7164/200yes
i1-IQ3_M15.884.078.3–88.4168/200yes16 GB pick
IQ4_XS-exps19.387.081.6–91.0174/200yes24 GB pick
Q4_K_M21.785.079.4–89.3170/200yes

Sage-A1.5: diff from the original (KLD)

No KLD for this model: The BF16 reference (about 72 GB) does not fit in our 32 GB test Mac (minipro), so there is no KLD for this model; we measured accuracy, tool calls and speed only.

Sage-A1.5: generation speed by quant and machine

Longer bar = faster replies. Smaller files are usually faster; the fastest measured here is Q4_K_M (InternScience) at 51 tok/s on Apple M2 Pro 32GB (llama.cpp). Compare bars of the same colour only: each colour is a different machine.

Sage-A1.5: generation speed by quant and machineHorizontal bars of measured generation speed in tokens per second for 4 files on 1 machine setup(s): Apple M2 Pro 32GB (llama.cpp). Longer = faster. Data table follows.020406080Generation speed (tokens per second), higher = fasterIQ4_XS-expsna0x2c6 · 19.3 GBIQ4_XS-exps by na0x2c6 · 19.3 GB · 50.1 tok/s on Apple M2 Pro 32GB (llama.cpp)50Q4_K_MInternScience · 21.7 GBQ4_K_M by InternScience · 21.7 GB · 50.9 tok/s on Apple M2 Pro 32GB (llama.cpp)51i1-IQ3_Mmradermacher · 15.8 GBi1-IQ3_M by mradermacher · 15.8 GB · 50.2 tok/s on Apple M2 Pro 32GB (llama.cpp)50i1-Q2_Kmradermacher · 13.2 GBi1-Q2_K by mradermacher · 13.2 GB · 49.2 tok/s on Apple M2 Pro 32GB (llama.cpp)49
  • Apple M2 Pro 32GB (llama.cpp)

Across: (tokens per second)

Show the numbers (table)
FilePublisherGBMachinetok/sPlacement
IQ4_XS-expsna0x2c619.3Apple M2 Pro 32GB (llama.cpp)50.1full GPU
Q4_K_MInternScience21.7Apple M2 Pro 32GB (llama.cpp)50.9full GPU
i1-IQ3_Mmradermacher15.8Apple M2 Pro 32GB (llama.cpp)50.2full GPU
i1-Q2_Kmradermacher13.2Apple M2 Pro 32GB (llama.cpp)49.2full GPU

Not sure which fits? Use the quant picker →

Recommendation by VRAM

"Best" = highest measured accuracy (500 items pooled) among files that fit fully on the GPU (file + 8192-token KV cache), tool-call score as tie-breaker. No KLD for this model (reference too large).

8GB GPU

Best accuracy that fits

Nothing we've tested fits fully in 8GB at 8192 context. Try a smaller model, run it GPU+RAM split, or rent a bigger GPU.

No GPU big enough? Rent one on RunPod or Vast.ai.

Referral links: QuantDiff may earn credit if you sign up. Referral partners have no say in scores or rankings.

12GB GPU

Best accuracy that fits

Nothing we've tested fits fully in 12GB at 8192 context. Try a smaller model, run it GPU+RAM split, or rent a bigger GPU.

No GPU big enough? Rent one on RunPod or Vast.ai.

Referral links: QuantDiff may earn credit if you sign up. Referral partners have no say in scores or rankings.

16GB GPU

Best accuracy that fits
i1-IQ3_M by mradermacher
File size
15.8 GB
no KLD: reference too large
Accuracy (500 items)
88.0% (84.9–90.6)
84.0%
Accuracy
Generation ()
50 t/s on Apple M2 Pro 32GB full GPU
Sage-A1.5.i1-IQ3_M.gguf
How to run this file

From mradermacher/Sage-A1.5-i1-GGUF, file Sage-A1.5.i1-IQ3_M.gguf.

(downloads it, then starts a chat; type /bye to quit):

ollama run hf.co/mradermacher/Sage-A1.5-i1-GGUF:Sage-A1.5.i1-IQ3_M.gguf

: open the Discover tab, search mradermacher/Sage-A1.5-i1-GGUF, pick the i1-IQ3_M download, then load it from the Chat tab.

Step-by-step guide

All Sage-A1.5 quants

QuantPublisherSizeFits meanKLD p99PP Gen Run
i1-Q2_K mradermacher imatrix 13.2 GB 16GB+ n/a n/a — 82.0% 87.3 ≈
81–92		81.0 ≈
75–86		94.0 ≈
89–97		 626 49 Apple M2 Pro 32GB settings & logs
i1-IQ3_M mradermacher imatrix 15.8 GB 16GB+ n/a n/a — 84.0% 92.0 ≈
87–95		79.5 ≈
73–85		95.3 ≈
91–98		 585 50 Apple M2 Pro 32GB settings & logs
IQ4_XS-exps na0x2c6 imatrix 19.3 GB 24GB+ n/a n/a — 87.0% 94.7
90–97		82.5
77–87		96.0
92–98		 614 50 Apple M2 Pro 32GB settings & logs
Q4_K_M InternScience 21.7 GB 24GB+ n/a n/a — 85.0% 94.0 ≈
89–97		82.0 ≈
76–87		96.0
92–98		 612 51 Apple M2 Pro 32GB settings & logs

Accuracy cells: % correct, with the 95% confidence interval (Wilson) underneath. ≈ = statistically tied with the best quant of the same model (intervals overlap). ▼ below = clearly worse than that model's best quant (intervals do not overlap). Ties are not ranked: sorting by an accuracy column puts tied quants together and orders them by KLD. How accuracy is measured.

File sizes are in GB as shown on Hugging Face (1 GB = 1 billion bytes). Measured on: tool calls: minipro (Apple M2 Pro 32GB); speed: minipro (Apple M2 Pro 32GB); accuracy: minipro (Metal, Apple M2 Pro (Mac mini), 32 GB unified memory). On a different GPU, expect generation speed to scale roughly with memory bandwidth. Task accuracy (GSM8K / MMLU / ARC-C, % correct on fixed subsets of 150 / 200 / 150 items, frozen suite quantdiff-lmeval-v1) was measured with the same GGUF files and llama.cpp build on Ghost (CUDA, NVIDIA GeForce RTX 5070 12 GB (64 GB RAM)), minipro (Metal, Apple M2 Pro (Mac mini), 12-core CPU, 19-core GPU, 32 GB unified memory), minipro (Metal, Apple M2 Pro (Mac mini), 32 GB unified memory). Accuracy depends on the weights, not on the GPU, but small numeric differences between backends can flip individual answers. With these sample sizes, differences under about 5 points are within noise (± shown on run pages). Methodology.

GPU too small? Rent one.

Run a 32B model at Q6/Q8, or a 70B, on a cloud GPU for an hour or two instead of buying hardware.

RunPod

Cloud GPU pods with 24-80GB cards (RTX 4090 / 5090 / A100 / H100) for big quants or BF16 reference runs.

Rent on RunPod
Vast.ai

GPU marketplace for renting a 24-48GB card for an evening of testing.

Rent on Vast.ai
Your GPU cloud here

Sponsor slot. Sponsors never influence scores or rankings.

Sponsor QuantDiff

Referral links: QuantDiff may earn credit if you sign up. Referral partners have no say in scores or rankings.