Qwen3.8-27B-Splash (mirror)

⚠️ Settings: read this before you run it

Setting	Use this	Why
temperature	0.8	What I settled on for Qwen3.8-27B after it looped at Qwen3.6's settings.
top_p	0.9	Same tuning. Honored by the runtime.
top_k	32	Honored. Must be an integer from 1 to 32. I'd tuned 50 on other runtimes, but here anything above 32, or "off", returns HTTP 400: Splash top-k must be an integer from 1 to 32.
max_tokens	cap it, ~20k	This is your only loop guard. The Splash runtime has no working repetition penalty, so a runaway turn only ends at the cap or the timeout.
repeat_penalty, presence_penalty, frequency_penalty	don't rely on them	The Splash runtime ignores all three. I verified this on the 3.6 build, and it's the same runtime.
Thinking	on by default	Short answers stay short: "say hi in three words" took 117 tokens here, against 2,815 on the 3.6 build.

LM Studio per-model default (saved as ~/.lmstudio/.internal/user-concrete-model-default-config/incoai/Qwen3.8-27B-Splash.json, which LM Studio reads live):

{
  "preset": "",
  "operation": {
    "fields": [
      { "key": "llm.prediction.temperature", "value": 0.8 },
      { "key": "llm.prediction.topPSampling", "value": { "checked": true, "value": 0.9 } },
      { "key": "llm.prediction.topKSampling", "value": 32 },
      { "key": "llm.prediction.contextOverflowPolicy", "value": "rollingWindow" }
    ]
  },
  "load": { "fields": [] }
}

---

A byte-for-byte mirror of incoai/Qwen3.8-27B-Splash, kept here as a pinned copy for my local agent stack. All credit for the packing, the DFlash 2 draft, and the Splash engine goes to Inco AI. Read their card first. This one only adds what I found running the Splash runtime.

Verified identical to upstream on 2026-09-23: the artifact_set_sha256 in manifest.json matches incoai's published manifest (b0ec3657…8da81).

What's in the package

It's not a Transformers or MLX checkpoint. It only loads in Splash, or in LM Studio's Splash runtime.

Path	Contents
target/	Qwen3.8-27B dense, 4-bit packed (splash-packed-q4)
draft/	DFlash 2 draft model, 5 layers, proposes 7 tokens per step
vision/	Vision encoder
tokenizer/	Tokenizer and chat template
manifest.json	Artifact hashes, execution geometry, upstream revisions

Upstream sources, from the manifest: target, tokenizer, and vision from mlx-community/Qwen3.8-27B-4bit @ 3e6447f, draft from incoai/Qwen3.8-27B-DFlash2 @ dedf8df. About 17.4 GB on disk.

Running it

LM Studio: it shows up as qwen3.8-27b-splash and reports model format yuzu.

Splash: upstream's command is splash serve --model incoai/Qwen3.8-27B-Splash. I haven't tried pointing Splash at this mirror's repo id, so use upstream's if you're going that route.

Measured

LM Studio on an M5 Max (128 GB), OpenAI-compatible endpoint, non-streaming, at the settings above. Timings include prefill:

Prompt	Tokens	Time	tok/s
Say hi in three words	117	1.4 s	82
Capital of France	33	0.3 s	100
Haiku about rain	208	2.1 s	99
300-word story	3,636	35.4 s	103

So it runs at about 100 tok/s. The Qwen3.8-27B MTPLX build did about 24 tok/s in the same agent. It's not as fast as the 3.6 MoE (about 148 tok/s), but it's far less wordy in its reasoning, and in my agent that matters more.

Sampling notes (LM Studio's Splash runtime, 2026-09-23)

I tested these on the 3.6 sibling, not on this model. It's the same runtime, so I'd expect the same behavior, but that's unverified:

• Honored: temperature, top_p, top_k, plus LM Studio's per-model config defaults.
• Ignored: repeat_penalty, presence_penalty, frequency_penalty.

With no working repetition penalty, the 3.6 build fell into a reasoning loop on its first evening (about 112k tokens of one repeated line, and no answer). If you run this behind an agent, cap max_tokens. Also, avoid greedy decoding, because Qwen's guidance is that it makes endless repetition more likely.

If you drive LM Studio from code

@lmstudio/sdk 1.5.0 hangs on llm.listLoaded() whenever a Splash model is loaded (it doesn't know the yuzu format). 2.0.0 fixes it.

License

Apache 2.0, same as upstream and the Qwen base model.