# MiniCPM vs Qwen3.8 Flash Next: Four-Turn Growing-Context Measurements

- **Canonical URL:** https://www.probackend.com/ai-local-model-deployment/how-two-local-moe-models-handled-a-growing-131k-context-session
- **Section:** News & Analysis
- **Published:** 2026-11-17
- **Summary:** Compare measured MiniCPM and Qwen3.8 Flash local deployment latency and memory on one M4 Max; 131K is configured capacity, not prompt length.

---

## Four turns, one growing context

In this four-turn, single-request probe, MiniCPM-V’s measured time to first token (TTFT) was 21.4 seconds at roughly 30K tokens, 36.6 seconds at roughly 62K, and 103.2 seconds at roughly 118K. Its cached follow-up took 3.3 seconds. Qwen3.8 Flash Next measured 76.6, 94.9, and 176.4 seconds at those stages, then 0.5 seconds for the cached follow-up.[^runs] These are observations from these runs, not general performance claims.

Both models met the 240-second first-token readiness threshold in this probe. MiniCPM-V had lower TTFT on the first three turns; Qwen3.8 Flash Next had the faster cached follow-up. Neither model is an overall winner.[^runs]

The configured context capacity was 131K, but the largest prompt in this sequence was approximately 118K. Because the sequence reused a growing prefix, these results do not establish cold-request performance at the same length or at the configured maximum.[^runs]

**Setup.** Apple Silicon workstation (M4 Max, 128 GB unified memory); mlx-serve v26.10.2. Bolt was benchmarked using a custom loader build and then dropped from the stack. Models: [pgmharikrishnan/MiniCPM-V-4.7-35B-A3B-MLX-Serve-4bit](https://huggingface.co/pgmharikrishnan/MiniCPM-V-4.7-35B-A3B-MLX-Serve-4bit), [mlx-community/KAT-Coder-V2.5-Dev-OptiQ-4bit](https://huggingface.co/mlx-community/KAT-Coder-V2.5-Dev-OptiQ-4bit), [NVIDIA Nemotron-3.5 Bolt 74B-A9B 4-bit experimental conversion](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16), and [ddalcu/Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-4.7bpw](https://huggingface.co/ddalcu/Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-4.7bpw). Tests were conducted in October 2026.[^runs]

## Reported comparison

| Measure | MiniCPM-V | Qwen3.8 Flash Next iQ |
|---|---:|---:|
| MMLU-Pro | 91.0% | 91.5% |
| GSM8K | 98% | 98% |
| Coding checks | 6/6 | 6/6 |
| Peak memory | 23.2 GiB | 77.9 GiB |
| Growing-session TTFT, turns 1/2/3/follow-up | 21.4 / 36.6 / 103.2 / 3.3 s | 76.6 / 94.9 / 176.4 / 0.5 s |[^runs]

In these reported checks, Flash had the higher MMLU-Pro score; GSM8K and the coding checks tied. The checks are limited and do not establish overall quality or reliability. The peak-memory figures describe these runs, not universal hardware requirements.[^runs]

The 240-second budget was the readiness threshold used for this probe. Meeting it in these runs does not establish that another request will meet it. In particular, cached-prefix results do not establish cold-request timing or success at the same prompt length.[^runs]

## Short decoding and cache observations

In a short plain-decode sample with MTP off, Flash measured 29.985 tokens per second. Its best MTP-on sample measured 42.138 tokens per second, at about 2.4 tokens per step. These samples were short—roughly 60 tokens or fewer—and are observations from this test, not stable throughput benchmarks. MTP showed little help at depth in this probe.[^runs]

MiniCPM-V’s short-prompt prefill measurements were noisy and are omitted. Flash’s mixed 4–8-bit variant also showed a wide third-turn spread in these runs, from 138.5 to 273.1 seconds. Neither observation supports treating one measured latency as repeatable behavior.[^runs]

The growing-session probe used cached-prefix conditions; Flash used MTP. Those conditions distinguish this test from four cold prompts and from a cold prompt at the configured maximum context.[^runs]

## What these runs show

On this machine and in this probe, MiniCPM-V used less measured peak memory and had lower TTFT on the first three growing turns. Flash had the faster cached follow-up and a higher best short MTP-on decode sample than its plain-decode sample. These results describe different measured dimensions, not an overall ranking.[^runs]

For related measurements, see the published [MiniCPM-V 4.7 35B context benchmark](/articles/performance-profiling-of-minicpm-v-4-7-35b-a3b-131k-context-on-apple-silicon-m4), [Qwen Flash Next growing-session probe](/articles/qwen-flash-next-on-m4-max-cached-long-prefix-timing-and-short-decode-results), and [four-model growing-context overview](/articles/four-local-moe-models-under-a-growing-context-load-m4-max-measurements). The comparison article’s canonical target is `/ai-models/minicpm-35b-vs-qwen38-flash-next-131k-context`.

**Disclaimer.** These measurements come from one machine, one server build, and one test suite on the stated date. Results vary with hardware, software version, settings, and load. They are not a general ranking or a recommendation.[^runs]

[^runs]: Supplied benchmark-run details for this article. The reported figures and test conditions are observations from those runs; no broader performance claim is implied.

![four turns, one growing context](https://assets.probackend.com/b8f0464b-d7e9-425e-bc49-f1c780bdf1fc?token=ast_7b9ee8975a4b187969d396286e3f9717c42f2331)

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1200 630" role="img" aria-label="Abstract illustration of a growing context and reused prefix"><defs><linearGradient id="bg" x1="0" y1="0" x2="1" y2="1"><stop offset="0" stop-color="#101a32"/><stop offset="1" stop-color="#263d65"/></linearGradient><linearGradient id="flow" x1="0" y1="0" x2="1" y2="0"><stop offset="0" stop-color="#52d6c7"/><stop offset="1" stop-color="#a4a9ff"/></linearGradient></defs><rect width="1200" height="630" fill="url(#bg)"/><path d="M115 390 C270 390 295 240 465 240 S650 390 795 390 S960 210 1090 210" fill="none" stroke="url(#flow)" stroke-width="8" stroke-linecap="round" opacity=".9"/><g fill="#52d6c7"><circle cx="115" cy="390" r="13"/><circle cx="465" cy="240" r="13"/><circle cx="795" cy="390" r="13"/><circle cx="1090" cy="210" r="13"/></g><g fill="none" stroke="#8199c3" stroke-width="2" opacity=".42"><rect x="95" y="430" width="180" height="78" rx="18"/><rect x="305" y="430" width="180" height="78" rx="18"/><rect x="515" y="430" width="180" height="78" rx="18"/><rect x="725" y="430" width="180" height="78" rx="18"/><rect x="935" y="430" width="180" height="78" rx="18"/></g><path d="M115 469h130M325 469h130M535 469h130M745 469h130M955 469h130" stroke="#a4a9ff" stroke-width="5" stroke-linecap="round" opacity=".55"/><path d="M115 530 C330 580 850 580 1090 530" fill="none" stroke="#52d6c7" stroke-width="3" stroke-dasharray="7 12" opacity=".5"/></svg>
