Splash 1.4.0: 120 tok/s on 64GB M5 Pro, 3x Faster Than Ollama

On October 21, Jian Chen posted on X that Splash 1.4.0 has been officially released.

In one sentence: a 64 GB M5 Pro running Qwen3.8-Flash-Next hits 120 tok/s decode and 1.3K tok/s prefill @32K context , beating the rival Strata by over 30% on both prefill and decode.

Data Recap: What the 64 GB M5 Pro Achieved

Decode up to 120 tok/s : Qwen3.8-Flash-Next on M5 Pro (64 GB).

Prefill 1.3K tok/s @ 32K : throughput when feeding long documents or codebases.

Over 30% faster than Strata : both prefill and decode lead.

Up to 1M context , and "larger models fit into smaller memory".

Why does 64 GB hold Qwen3.8-Flash-Next? Public specs show it is a 125B total-parameter, 6B-active MoE model with native 1M context support. After quantization the weights fit exactly into 64 GB unified memory — "larger model into smaller memory" is the result of quantization plus MoE optimization combined.

What Is Splash: Built Around Model and Apple Silicon

Splash is Inco AI's open-source inference engine ( github.com/incoai/splash), positioned as "built around the model and around Apple silicon" .

The September 19 debut showed Qwen3.8-27B on M5 Max MacBook Pro at 144 tok/s , with three multipliers:

Decode up to 3× Ollama ;

Up to 2× oMLX ;

Agent fan-out (splitting into multiple sub-agents) near 4× .

Version 1.4.0 extends that "speed" to a more mainstream machine — a 64 GB M5 Pro can now enjoy it.

Interpreting the Numbers (Including a Data Trap)

① 6B active parameters are the physical basis for speed. MoE moves only 6B parameters per token, so the M5 Pro's memory bandwidth can keep up — this is why a 125B model reaches 120 tok/s on 64 GB. It matches the bandwidth-wall formula: smaller active parameters → higher tok/s.

② What 120 tok/s means for agents. The previous DeepSeek run at 63.8 tok/s built a game world in 65.7 minutes; 120 tok/s roughly halves that time, making the "waiting feeling" in interactive agents virtually disappear.

③ ⚠️ "2× oMLX" — do not compare across conditions. This is the official same-condition benchmark. You cannot take "M5 Pro running a 6B-active model" speed and directly compare it with "M5 Ultra 256 GB running a 24B-active model" at 63.8 tok/s — the per-token data movement differs by 4× and the hardware tier is different. Cross-model, cross-machine speed comparisons must first check test conditions.

④ Agent fan-out near 4× is the most noteworthy. Single-stream 3× is a "benchmark"; 4× concurrent sub-agent speedup is "production" — today's real speed-hungry workloads are multi-agent parallel scenarios, exactly what Splash optimizes for.

What This Means for You

64 GB Macs (M5 Pro/Max) can now run high-speed large-model inference — no Ultra required.

Still on Ollama? Worth a try : same machine, same model, official data shows up to 3× gap.

Running agent workflows : fan-out optimization hits the bullseye — the more sub-agents, the larger the gap.

Open-source address: github.com/incoai/splash Next up I plan a same-machine live test: Splash vs oMLX vs Ollama on the same model to verify the "3×" and "2×" claims. If you want to see it, click "Read Later"; I'll schedule by popularity.

(Data sources: Jian Chen @jianchen1799 2026-10-21 tweet, Inco AI @inco_ai 2026-09-19 tweet; Qwen3.8-Flash-Next specs from public info. Speed multipliers are official figures, not independently reproduced.)