steadfastgaze/Qwen3.8-Flash-Next-MoEspressoV3

Qwen3.8-Flash-Next for MoEspresso

This package brings Qwen3.8-Flash-Next to MoEspresso. Qwen3.8-Flash-Next is a 125B Mixture-of-Experts model with 6B active parameters per token and a 51B-parameter PLE/n-gram memory.

On a 2021 M1 Max with a 24-core GPU and 32 GB of unified memory, MoEspresso serves this package at roughly 12–15 decode tokens/s. The local bounded configuration also scored 84.3% on a frozen 48-question generated-answer comparison, ahead of Claude Opus 4.8 xhigh at 80.3% on the same questions.

The complete download occupies 145.75 GB (135.74 GiB), including the original BF16 PLE tables. MoEspresso keeps only its active working set in unified memory.

Run it

The package requires MoEspresso 3.0.0 on an arm64 Apple Silicon Mac running macOS 26.2 or later. Its files use the MoEspresso package format and cannot be loaded directly by Transformers, mlx-lm, or GGUF runtimes.

brew install steadfastgaze/tap/moespresso
brew install hf

hf download steadfastgaze/Qwen3.8-Flash-Next-MoEspressoV3 \
  --local-dir ./models/qwen3.8-flash-next

moespresso verify ./models/qwen3.8-flash-next
moespresso serve ./models/qwen3.8-flash-next

The server exposes POST /v1/chat/completions and GET /health at http://127.0.0.1:8080. Thinking is enabled at medium effort by default, and requests may select low, medium, or xhigh. Ordinary serving uses a 128K context limit. --max-context-tokens can select a supported limit up to the package's 262,144-token architecture limit, with larger limits reserving more cache memory and leaving less room for resident experts.

With the server running, measure decode speed in another terminal:

moespresso speed

For one-shot generation:

moespresso generate ./models/qwen3.8-flash-next \
  --prompt "Write a lock-free bounded queue in C++ and explain its memory ordering." \
  --max-tokens 2048

Quality on the bounded path

The comparison uses a frozen set of 48 generated-answer questions, with eight questions from each of six public LiveBench categories: coding, data analysis, instruction following, language, math, and reasoning. A fixed salted SHA-256 rule derives the selection from pinned datasets, and the official task scorers grade every complete answer. The benchmark reproduction kit contains the selection procedure, runner, isolated graders, and reports.

Model / configuration	Six-category macro
Qwen3.8 Flash, xhigh	90.7%
GPT-6 Sol, medium	89.2%
Claude Opus 5.5, low	87.6%
MoEspresso, IQ_K routed experts, Cache-Prior 2/2, capacity 223, medium	84.3%
GPT-6 Luna, xhigh	81.4%
Claude Opus 4.8, xhigh	80.3%

The local row used medium thinking, greedy generation, 223 resident expert slots per layer, and the default Cache-Prior policy with factor 2 and two protected routes. Disk KV was disabled. All 48 answers stopped naturally, with no output-cap hits. All hosted rows ran through OpenRouter. Hosted Qwen used fixed sampling controls, while the GPT and Claude endpoints used their provider-native defaults.

This is a custom comparison derived from public LiveBench tasks. It measures the actual bounded route used for 32 GB serving. The official LiveBench leaderboard uses a different generation protocol.

SSD-backed experts and Cache-Prior routing

Dense weights and shared experts remain in unified memory. MoEspresso assigns the remaining memory budget to a separate routed-expert pool for each layer, then loads a missing expert row from SSD when the router selects it.

Bounded serving uses a policy inspired by Cache-Prior. During ordinary single-token decode from layer 2 onward, the router's two strongest original choices remain protected. Resident candidates receive a factor-two ranking preference for the remaining positions. Ten routed experts still execute, and their contribution weights come from the original router probabilities normalized over the selected set.

Prefill is unbiased. Layers 0 and 1 also retain the model's original routing during decode, as does fully resident execution. Every expert remains available, and a selected nonresident expert is loaded from SSD. Because the resident preference can change the selected expert set and the model's output, the quality result above was measured with this policy enabled.

The Cache-Prior documentation explains the full algorithm and its relationship to REAP.

Package recipe

The model has 48 layers. Each MoE layer retains all 512 routed experts and selects ten per token alongside one shared expert. The routed experts' gate, up, and down projections use IQ3_K in layers 0 and 1, covering six projection groups. The remaining 138 groups use IQ2_K. Dense and non-routed tensors use Q6_K, Q8_0, or BF16 according to their role.

The original BF16 PLE/n-gram tables stay on SSD, where MoEspresso reads only the selected rows instead of placing the complete 95.37 GiB payload in unified memory. The Qwen attention cache stores its older body in KVarN K4/V4 while keeping the sink and recent suffix exactly in BF16. This text-only package excludes the vision encoder and MTP sidecar.

Payload	GB	GiB
Model safetensors	43.35	40.37
Original BF16 PLE/n-gram tables	102.40	95.37
Complete declared payload	145.75	135.74

The package was derived from Qwen/Qwen3.8-Flash-Next at revision de4b8e4d43b917e7706784d8bb445c9af86a3540.

Download verification

moespresso verify checks the package manifest, declared paths, byte sizes, SHA-256 digests, tensor keys, and generated compatibility files. The package manifest identity is pkg:e469be1b1a19e966d4f546c8eaa7b5ee980ad19f0e41883d3c25678a4728fef5.

License and credits

The model weights are released under the Qwen Community License 1.0, which includes separate terms for commercial Model-as-a-Service and AI Work Assistant use.

The Qwen team trained and released the base model. The IQ_K formats were designed by Iwan Kawrakow. MoEspresso executes the IQ_K tensors through mlx-iqk. The K-quant tensors run through the project's mlx-kquant fork, which derives from Asher Feldman's original integration. MoEspresso runs on Apple's MLX.