AI
AI Updates
Back to all posts
·
ai-updatesdaily-briefing

Qwen3.8 Flash Drops, llama.cpp 0.3.0 Fixes RDNA3, Jalapeño Chip Rumors

Alibaba releases Qwen3.8-Flash variants, llama.cpp 0.3.0 fixes AMD GPU hangs, and OpenAI's Jalapeño chip benchmarks dominate the HN discussion.

A
AI Updates

🔥 Top Stories

1. Qwen3.8-Flash-Next Arrives (FP8 & Next Gen)

Alibaba just dropped the heavy lifting onto the community. The official Qwen3.8-Flash-Next and its FP8 variant are live on HuggingFace. For self-hosters, this is the “fastest” version of the Qwen3.8 family. The 27B model is now available in a quantized format that likely cuts VRAM requirements by 30% compared to FP16, making it viable for high-end consumer GPUs.

  • The Shift: Previously, running 27B models required 16GB+ VRAM at Q4. The Flash variant suggests architectural optimizations for faster attention.
  • VRAM Fit: The 27B FP8 version is estimated to run on RTX 3090 24GB ✓ or RTX 4090 24GB ✓.
  • Why Mention: This brings a top-tier Chinese model into the fold for local inference without needing a data center GPU.
  • HuggingFace

2. DeepSeek-V4-Pro Surges Against R1

While DeepSeek-R1 holds the #2 spot on HF trending (13,589 likes), the new DeepSeek-V4-Pro is gaining rapid traction (#12 on HF trending with 5,482 likes). It appears to be the efficiency-focused counterpart to the reasoning-heavy R1.

  • The Shift: R1 is the “brain,” V4-Pro seems to be the “engine.” Self-hosters looking for high-throughput inference over pure reasoning complexity should watch this one.
  • VRAM Fit: Assuming similar architecture to R1 (likely 67B or 256K context), the Q4_K_M quantization likely requires 18-22GB VRAM. Fits RTX 3090/4090 ✓ but needs 2x 3090s for full context.
  • Source: HuggingFace Trending

3. OpenAI Jalapeño: “Better Than Blackwell”

The biggest hardware talk this week isn’t about GPUs—it’s about the new OpenAI “Jalapeño” chip. A recent newsletter analysis claims this custom silicon outperforms Nvidia’s Blackwell B200 in specific AI inference tasks.

  • The Impact: For the self-hosting community, this signals a shift in the hardware arms race. If OpenAI chips are more efficient for inference, the price/performance gap for consumer GPUs might widen in favor of cloud inference for heavy models (120B+).
  • Context: HN Thread (530 points)

📊 Model Trends

HuggingFace Trending: The Self-Hosting Shift The top models are a mix of established image models and new LLM contenders. Note the dominance of the 27B class in recent trending lists.

Model Likes Category VRAM (Q4 est.) Hardware Fit
black-forest-labs/FLUX.1-dev 14,256 Text-to-Image ~16GB RTX 3090 24GB ✓
deepseek-ai/DeepSeek-R1 13,589 LLM (Reasoning) ~24GB RTX 4090 24GB ✓
Qwen/Qwen3.8-27B 12,820 LLM (General) ~18GB RTX 3090 24GB ✓
moonshotai/Kimi-K3 11,005 LLM (Long Context) ~22GB RTX 4090 24GB ✓
stabilityai/stable-diffusion-xl 8,074 Image Gen ~10GB RTX 3060 12GB ✓
  • Observation: FLUX.1-dev is dominating the image generation space, while Qwen3.8 is capturing the LLM market. The 8B models (Llama 3.1) are holding steady but losing traction to the 27B class for complex tasks.
  • Source: HuggingFace Trending

⚙️ Engine Updates

Crucial for AMD Users: llama.cpp b10636 The latest llama.cpp build (b10636) includes a critical fix for RDNA3 integrated GPUs. Previous builds suffered from hangs due to GDN cache fusion on HIP.

  • The Fix: PR #27722 (cuda: disable GDN cache fusion on HIP — hangs RDNA3 iGPUs) ensures that users with AMD APUs or RDNA3 laptops won’t crash during inference.
  • Why It Matters: If you self-host on a laptop or use AMD hardware, this is a stability update.
  • GitHub PR

vLLM 0.28.0: Performance Push The latest vLLM release focuses on throughput. While specific benchmarks aren’t detailed in the raw data, the 0.28.0 tag suggests significant optimization for agentic workloads (linked to DigitalOcean’s inference engine news).

  • Impact: Better for serving multiple users. If you run vLLM for production, update immediately.
  • GitHub Release

Ollama v0.33.0: Easier Deployment Ollama v0.33.0 has landed, likely incorporating backend improvements from vLLM and llama.cpp.

  • Action: Check your ollama serve logs for new model support (e.g., Qwen Flash).
  • GitHub Release

📰 AI News

1. Z.ai Confirms Ox Alpha (GLM Series)

Z.ai has officially confirmed Ox Alpha, a new GLM-series model that rivals DeepSeek in performance. This adds to the competitive pressure on the Llama 3 and Qwen families.

  • Context: Bloomberg (215 points on HN).
  • Self-Hoster Take: Expect weight releases soon. If Ox Alpha matches DeepSeek’s efficiency, it could become a top-tier model for 27B class deployment.

2. Bill Gates: “The Turbulent AI Era is Here”

Bill Gates published a note on AI governance and the “turbulent AI era.”

  • Context: Gates Notes.
  • Takeaway: While not technical, this signals a regulatory shift that could impact open-weight model hosting in certain jurisdictions.

3. Fake Thinktank Scandal

A fake US thinktank funded by Israel was shut down for gaming AI for propaganda.

  • Context: The Guardian.
  • Takeaway: Highlights the need for robust model filtering in enterprise environments.

🔄 What Changed Since Last Scan

New Releases:

  • Qwen3.8-Flash-Next: Official release of the Flash variant (3044 likes).
  • Qwen3.8-Flash-Next-FP8: FP8 quantization for lower memory usage (29 likes).
  • llama.cpp b10632 - b10636: Includes the RDNA3 fix and Nemotron 3.5 Lightning layer fixes.
  • vllm v0.28.0: Latest inference engine release.
  • ollama v0.33.0: Major version bump for Ollama.

Notable Community Quantizations:

  • Qwen3.8-27B-Uncensored-GGUF: (JonathanColetti) High download count indicates strong demand for uncensored variants.
  • Qwen3.8-27B-OBLITERATED: (OBLITERATUS) Aggressive quantization for 12GB VRAM.
  • HuggingFace

Sources & Links