🤯 A llama.cpp tweak made Qwen3.8-Flash-Next 2–3× faster at prompt processing on a DGX Spark.
And the optimization is basically is to optimize SSD reading. Seems almost trivial 😂
Since the Qwen3.8-Flash-Next has a HUGE PLE/n-gram table that can live on SSD instead of consuming precious RAM/VRAM.
However, llama.cpp's current mmap path can trigger thousands of tiny scattered page faults, often pulling 4KB when the model may only need ~90 bytes.
PR #28136 introduces a new direct-read path using parallel pread() workers.
Reported DGX Spark results:
🐢 Before → ~300 tok/s prefill
🚀 After → ~750–800 tok/s
Cold diverse-text tests show roughly:
2.0–3.1× faster prompt processing
WITHOUT loading the giant PLE table into RAM.
🧠 Strix Halo 128GB testers are reporting +20–32% cold prefill too.
⚠️ PR #28136 is currently OPEN and not merged yet.
🔗 llama.cpp PR #28136 👇
As we prepare to release Astra, we’re focused on making increasingly capable AI safe and broadly accessible.
Astra represents a significant advance in cybersecurity capability, reaching the Critical threshold under our Preparedness Framework.
We're previewing how we evaluated the model, how its safeguards have advanced alongside its capabilities, and what we'll continue to learn and improve.
https://t.co/OrrTgdU90K
September very well may be the BIGGEST month in AI ever...
We're expecting:
(Sep 1–6)
-Fable 5.1 (rumored tomorrow)
-GPT 6 Astra (rumored Thursday)
-Grok 4.7
(Sep 7–15)
-DeepSeek V5
-Composer 3
-Gemini 3.8 flash
(Sep 16–22)
-Kimi K model (not Kimi K4)
-Potentially Opus 5.1
(Sep possible but who knows)
-MiniMax 2.7T open-weight model
-Meta Llama 5 / next Muse “Watermelon”
-Mistral frontier open MoE
Nobody knows exactly when all these will land but regardless it's going to be a INSANE month.
Which are you waiting for most?
DART: Detect Anything in Real Time
Detect Anything in Real Time: Real-time object detection using frontier object detection models.
Training-free framework that converts SAM3 into a real-time multi-class open-vocabulary detector. Achieves 55.8 AP on COCO val2017 (80 classes) at 15.8 FPS (4 classes, 1008px) on a single RTX 4080.
Today we're releasing Isaac 0.5: 36B dynamic MoE, open weight 🤗 embodied foundation model.
Isaac combines multimodal video understanding, embodied reasoning and robot control into a single, sparse backbone.
The VRAM barrier is officially dead.
I just ran Qwen 3.8 Flash Next (MoE) 125B A6B with a 250,000 context window on a single 24GB RTX 4090.
21 tokens/sec decode. 364 t/s prefill.
no mtp. no dflash. no kv cache quantization!
We are running datacenter models on consumer hardware.
Tested on Ubuntu 22 | CUDA 13.0 | PCIe 4.0 x16 | 110 GB DDR4 System RAM with a continuous 28k prompt across all runs.
### The Benchmarks & Scaling
# 1. Hybrid Offload (-ncmoe 40 @ 80k Context)
Offloaded 40 expert layers to the GPU, pushing VRAM to the ceiling.
./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 80000 --port 8080 -v --fit off -b 4096 -ub 4096 -ncmoe 40
Prefill: 383.85 t/s | Decode: 22.52 t/s
Footprint: 23.85 GB VRAM | 97 GB RAM
# 2. Full CPU MoE Offload (-cmoe @ 80k Context)
Pinned all 512 expert layers to DDR4 RAM (-cmoe), keeping attention on the 4090.
llama.cpp flags: (Same as above, replace -ncmoe 40 with -cmoe)
Prefill: 355.72 t/s | Decode: 20.84 t/s
Footprint: 11.66 GB VRAM (12GB+ VRAM freed up!) | 110 GB RAM
# 3. The 180,000 Context Run
Prefill: 357.75 t/s | Decode: 20.98 t/s | VRAM: 15.6 GB | RAM: 110 GB
# 4. The 250,000 Context Absolute Ceiling
./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 250000 --port 8080 -v --fit off -b 4096 -ub 4096 -cmoe
Prefill: 364.29 t/s | Decode: 20.97 t/s
Footprint: 18.3 GB VRAM (Still ~5.7 GB of VRAM headroom!) | 110 GB RAM
### Key Insights:
-b 4096 -ub 4096: doubles the prompt ingestion from ~150 to 364+ t/s.
-cmoe Free Lunch: Shifting expert layers to DDR4 RAM slashes VRAM from 24GB to 11.6GB with virtually zero decode penalty (22.5 -> 20.9 t/s), enabling the 250k context ceiling.
Qwen 3.8 Flash-Next (UD-Q4_K_XL) is a massive 111.4 GB model split across 4 shards. To run this architecture, you must build from the experimental PR branch (#27742) by @danielhanchen:
git clone && cd llama.cpp
git fetch origin pull/27742/head:qwen-next && git checkout qwen-next
cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native -DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release -j $(nproc) --target llama-server
A single 4090 paired with 100 GB of cheap DDR4 RAM will comfortably serve production grade 125B inference.
While Qwen 3.8 27B (dense) still holds the crown for single 3090/4090 rigs, Flash Next proves 125B hybrid models are officially viable on consumer hardware.
Hugging Face GGUF link and complete performance telemetry graphs are dropped in the replies below.
GLM 5.3 Flash VS Qwen 3.8 Flash Next, which one takes the open weights crown this week?
Introducing Index
Today we're coming out of stealth with Index, the largest & most diverse robot dataset in the world
→ 30min of video uploads/sec
→ 16M video uploads
→ Paid $15M to date
→ 264k downloads
We're committed to spending $1B the next 12 months on data & compute