Dataset Viewer
Auto-converted to Parquet Duplicate
model
stringlengths
27
52
params
stringlengths
2
11
memory_gb
float64
2.2
20.5
⌀
decode_tok_s
stringlengths
2
9
⌀
mode
stringclasses
4 values
basalt_version
stringclasses
2 values
hardware
stringclasses
1 value
notes
stringlengths
5
47
⌀
mlx-community/Qwen3-4B-4bit
4B
2.2
78-89
plain
0.8
M4 Pro 24GB
dense
Youssofal/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed
27B ternary
7.4
20.9-21.5
plain
0.8
M4 Pro 24GB
hybrid Gated DeltaNet
Youssofal/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed
27B ternary
7.4
27.1
--mtp (code)
0.8
M4 Pro 24GB
null
Youssofal/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed
27B ternary
7.4
31-38
--speculative-lookup (code)
0.8
M4 Pro 24GB
null
lukaskremla/Qwen3.8-27B-3bit-MLX-TextOnly
27B
null
15-19
plain
0.8
M4 Pro 24GB
hybrid
lukaskremla/Qwen3.8-27B-2bit-MLX-TextOnly
27B
null
18-22
plain
0.8
M4 Pro 24GB
loops on 3 of 8 prompts
mlx-community/gemma-4-e2b-it-4bit
2B
2.7
~110
plain
pre-0.8
M4 Pro 24GB
null
mlx-community/gemma-4-e4b-it-4bit
4B
4.3
~85
plain
pre-0.8
M4 Pro 24GB
null
mlx-community/Qwen3.5-4B-4bit
4B
2.3
~92
plain
pre-0.8
M4 Pro 24GB
null
mlx-community/Qwen2.5-Coder-7B-Instruct-4bit
7B
4.1
~60
plain
pre-0.8
M4 Pro 24GB
null
mlx-community/Qwen3.5-9B-OptiQ-4bit
9B
5.8
~37
plain
pre-0.8
M4 Pro 24GB
null
ornith-ai/Ornith-1.5-9B-MLX-4bit
9B
4.7
~48
plain
pre-0.8
M4 Pro 24GB
reasoning
LiquidAI/LFM2.5-8B-A1B-MLX-4bit
8B MoE
4.5
~76
plain
pre-0.8
M4 Pro 24GB
always reasons
LiquidAI/LFM2.5-24B-A2B-MLX-4bit
24B MoE
12.6
106.2
plain
0.8
M4 Pro 24GB
always reasons; measured 2026-11-05 via --stats
Ex0bit/MYTHOS-26B-A4B-PRISM-PRO-DQ-MLX
26B MoE
14.5
~14
plain
pre-0.8
M4 Pro 24GB
null
Ex0bit/Qwen3.6-35B-A3B-PRISM-MLX-NVFP4
35B MoE
20.5
~8
plain
pre-0.8
M4 Pro 24GB
NVFP4
mlx-community/gemma-4-26b-a4b-it-4bit
26B MoE
14.5
null
failed
pre-0.8
M4 Pro 24GB
gibberish output
mlx-community/Qwen3.6-27B-4bit
27B
14.5
null
failed
pre-0.8
M4 Pro 24GB
gibberish output
mlx-community/gemma-4-31b-it-4bit
31B
17
null
failed
pre-0.8
M4 Pro 24GB
OOM (needs 32GB+)

basalt on Apple Silicon: decode speed

Decode speed of MLX models served by basalt, a native Swift LLM server (OpenAI + Anthropic API + MCP, no Python at runtime).

  • Hardware: M4 Pro, 24 GB unified memory, on a 68 W adapter.
  • Release build, greedy decoding.
  • basalt_version = 0.8: measured November 2026. pre-0.8: earlier reports, not re-measured.
  • Rows with mode = failed are models that did not work (gibberish output, OOM). They are kept on purpose.

Only decode tok/s is published here. TTFT and comparisons against other runtimes are not measured yet.

Your Mac is not here? Run basalt fit, then open an issue with your chip, memory and tok/s.

Interactive picker: Bosalt/README Space.

Downloads last month
58