Spark-X3-MoE-10B-A1B -- Model Request

[residenthiago2011]

The Spark line currently ships dense 1.7B and 4B models. The next high-leverage release could be a
10B-total / 1B or 1.7B(SparkX2.5 size)-active Mixture-of-Experts model: roughly 2.5x the capacity of Spark-X2.6-4B,
but with decode cost close to Spark-X2.5-1.7B. It could fits in 6-7 GB at Q4 with a 32K context,
which keeps it inside flagship-phone RAM while giving laptop-class quality.

[FenjuFu]

Thanks for the proposal. It would help to separate total-weight storage, active-parameter compute and the KV cache at the intended context length: the RAM and decode estimates need measurements on the target runtime and hardware. Could you share the target phone/SoC or laptop, runtime and quantization format, the available memory budget, and representative tasks with latency/quality targets? That would make the trade-offs easier to assess. This discussion does not confirm a model implementation or release schedule.

[residenthiago2011]

On-Device LLM Comparison Research

Target Hardware

• Device: Poco F7 (model 25053PC47G), Snapdragon 8s Gen 4, Adreno 825 (OpenCL 3.0), 12 GB RAM
• Runtime: llama.cpp (build 972d231) → CPU-only used
• Quantization: Q4_K_M for both tested models

Models Under Test

Model	Type	Total params	Active params	Size (Q4_K_M)
LFM2.5-8B-A1B (LiquidAI)	MoE	8.3B	1.5B	4.79 GiB
Spark-X2.5-1.7B (XHToken)	Dense	1.7B	1.7B	1.03 GiB

Measured Inference (Poco F7, CPU, Q4_K_M)

Model	tg2048	pp4096	tg4096
LFM2.5-8B-A1B	20.74	58.20	21.99
Spark-X2.5-1.7B	20.05	74.48	19.12

RAM: total 12 GB. LFM footprint ~4.8 GB; Spark ~1.0 GB. Available memory during runs: LFM ~5.8 GB, Spark ~5.6–7.1 GB.

Published Intelligence Benchmarks (self-reported)

Shared benchmarks (reported by ≥2 models):

Benchmark	LFM2-8B-A1B	Spark-X2.5-1.7B	Spark-X2.5-4B	Mellum2.1-12B-A2.5B
IFEval	77.58	89.5	93.0	90.6
IFBench	25.85	66.3	75.0	—
GPQA	29.29	43.8	67.4	64.6 (Diamond)
AIME	—	69.4 (2026)	90.7 (2026)	83.3 (25/26)
SWE-bench Verified	—	28.3	41.6	47.0
SWE-bench Pro	—	10.4	44.4	28.0
BFCL v4	—	46.9	65.1	62.3
LiveCodeBench v6	21.04%	—	—	82.0
HumanEval+	69.51%	—	—	91.5
GSM8K/GSM-Plus	84.38	—	—	88.3
MMLU/MMLU-Redux	64.84	—	—	87.8

Key Points

1: Does more experts = more intelligence?

No. Quality tracks active parameters, not total or expert count:

• LFM2-8B-A1B (8.3B total, 1.5B active) is "comparable to 3-4B dense models" per its own card.
• Spark-X2.5-1.7B (1.7B active, dense) beats LFM on IFEval (89.5 vs 77.6) with fewer total params.
• Expert count buys capacity per FLOP, not automatic superiority: Mellum2.1 (64 experts, 8 active, 2.5B active) still trails Qwen3.5-9B (9B active) on GPQA Diamond (64.6 vs 77.8).

2: Trade-offs for on-device deployment

• Storage vs compute vs KV cache: MoE (LFM) stores 4.8 GB but computes like 1.5B dense; dense Spark stores 1.0 GB and computes 1.7B. At 4096 context both decode ~20 t/s, but Spark has 4.7× higher prompt throughput.
• Memory budget: 11.5 GB total; LFM leaves ~5.8 GB free, Spark ~5.6–7.1 GB. KV cache at 4096 context is modest.
• Quality: Spark leads agentic/instruction (IFEval 89.5, BFCL 46.9–65.1); LFM leads math (GSM8K 84.4). For coding, Mellum2.1 (12B/2.5B) dominates (LCB 82.0, HumanEval+ 91.5) but is 2.5B active — heavier.

Recommendations

• Best speed/quality for agentic tasks on Poco F7: Spark-X2.5-1.7B (dense, 1.0 GB, ~19–20 t/s gen, IFEval 89.5).
• Best math/reasoning per byte: LFM2.5-8B-A1B (MoE, 4.8 GB, ~22 t/s gen, GSM8K 84.4).

Open Items / Needed Measurements

• HTP (Hexagon) NPU — CPU-only numbers; OpenCL path may differ.
• Peak RSS per model at 4K/32K context not yet sampled.
• Quality evals on-device (not just published) not performed.
• Spark-X2.6-4B and Mellum2.1 tests
• Tests with ~10B Dense Models