AI
AI Updates

DeepSeek-V4-Pro Surges, FLUX Takes Over Image Generation

Qwen3.8-2.4T-A95B launches to immediate buzz; llama.cpp hits b10375 with Qwen optimizations; vLLM 0.27 brings Kimi K3 support; Grok 4.6 benchmarks drop.

A
AI Updates

🔥 Top Stories

1. Qwen3.8-2.4T-A95B: Alibaba’s Flagship Drops

Qwen/Qwen3.8-2.4T-A95B is the headline release of the day. Alibaba’s Qwen team unveiled a 2.4 trillion parameter Mixture-of-Experts model with 95 billion active parameters (A95B). The model immediately hit HuggingFace with 372 likes and climbed to #271 on Hacker News within hours.

The ecosystem is already moving: Community forks and quantized variants appeared within hours of the initial release:

Variant Likes Notes
Qwen/Qwen3.8-2.4T-A95B 372 Official flagship
Qwen/Qwen3.8-2.4T-A95B-FP8 93 Official FP8 quantization
unsloth/Qwen3.8-2.4T-A95B-GGUF 49 Unsloth GGUF conversion
huginnfork/Qwen3.8-27B-FP8 21 27B distilled variant, FP8
huginnfork/Qwen3.8-27B-NVFP4A16 10 27B distilled variant, NVFP4A16
Ma7ee7/Qwen3.8_4B_Distilled_GGUF 20 4B distilled GGUF

What about Qwen3.8-27B? The 27B distilled variant is already available through community forks — huginnfork/Qwen3.8-27B-FP8 (21 likes) and huginnfork/Qwen3.8-27B-NVFP4A16 (10 likes). An official release from the Qwen team may still follow, but community demand is clearly ahead of the official rollout. There’s also a 4B distilled variant available.

What this means for you: The flagship 2.4T model requires serious infrastructure (multi-GPU or cloud deployment), but the distilled variants make Qwen3.8 architecture accessible at consumer-friendly sizes. The 27B variant fits on an RTX 3090 at Q4_K_M. The 4B variant runs on even modest hardware.

2. Grok 4.6 Scores 61 on Artificial Analysis Intelligence Index

xAI’s Grok 4.6 landed on HN today (188 points, 169 comments) with benchmark results scoring 61 on the Artificial Analysis Intelligence Index. This represents a significant data point in the ongoing competitive landscape between xAI’s offerings and the broader open-weight ecosystem.

3. DeepSeek V4 Pro 0813 Update

DeepSeek V4 Pro 0813 topped today’s AI-related HN stories (444 points, 153 comments), signaling a fresh update to DeepSeek’s already-popular V4 Pro model. On HuggingFace, deepseek-ai/DeepSeek-V4-Pro holds 5,407 likes with 1.4M downloads — the third-most-downloaded LLM on the platform.

Rank Model Likes Key Tags
1 black-forest-labs/FLUX.1-dev 14,110 text-to-image, flux
2 deepseek-ai/DeepSeek-R1 13,568 deepseek_v3, reasoning
3 moonshotai/Kimi-K3 10,571 kimi_k3, compressed-tensors
4 stabilityai/SDXL 8,038 stable-diffusion
5 hexgrad/Kokoro-82M 6,679 text-to-speech
6 meta-llama/Meta-Llama-3-8B 6,628 llama
7 meta-llama/Llama-3.1-8B-Instruct 6,589 llama
8 openai/whisper-large-v3 6,138 whisper
9 black-forest-labs/FLUX.1-schnell 5,527 text-to-image
10 deepseek-ai/DeepSeek-V4-Pro 5,407 deepseek_v4
11 sentence-transformers/all-MiniLM-L6-v2 5,203 embeddings
12 Tongyi-MAI/Z-Image-Turbo 5,105 text-to-image
13 openai/gpt-oss-120b 5,100 gpt_oss
14 bigscience/bloom 5,035 bloom
15 stabilityai/stable-diffusion-3-medium 5,014 stable-diffusion

Notable entries: The leaderboard is dominated by a mix of reasoning models (DeepSeek-R1, Kimi-K3) and creative generation tools (FLUX.1-dev, Kokoro-82M). openai/gpt-oss-120b at #13 with 4M downloads shows strong sustained interest in OpenAI’s open-weight release. zai-org/GLM-5.2 (4,941 likes) and openai/gpt-oss-20b (4,908 likes) round out the top 20.

Qwen Model Family — Full Qwen3.8 Coverage

The Qwen family continues to dominate the open-weight landscape with three major version branches now active:

Qwen3.8 (Newest)

Model Likes Notes
Qwen/Qwen3.8-2.4T-A95B 372 Flagship — 2.4T MoE, 95B active
Qwen/Qwen3.8-2.4T-A95B-FP8 93 Official FP8
huginnfork/Qwen3.8-27B-FP8 21 Community 27B distilled
Ma7ee7/Qwen3.8_4B_Distilled_GGUF 20 Community 4B distilled

Qwen 3.8 27B — When Will It Drop?

The Qwen team has indicated the 27B variant is expected “any day now” — they said “next week” during last week’s announcements. This would put an official release in the mid-to-late August 2026 window.

This aligns with their pattern of releasing smaller, locally-runnable variants shortly after the flagship. The 27B model would be runnable on consumer GPUs (RTX 3090/4090 with quantization) — something the 2.4T flagship can’t offer for local inference.

Watch for: Qwen’s GitHub releases page and HuggingFace organization announcements for the official drop.

Qwen3.6

Model Likes Notes
HauhauCS/Qwen3.6-35B-A3B-Uncensored 3,404 Top community variant
Qwen/Qwen3.6-35B-A3B 2,668 Official 35B MoE
Qwen/Qwen3.6-27B 2,241 Official 27B dense
DavidAU/Qwen3.6-27B-Fable-Fusion 1,942 Community fine-tune

Qwen3.5

Model Likes Notes
Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled 2,938 Claude-opus reasoning distillation
Qwen/Qwen3.5-9B 1,807 9B dense
Qwen/Qwen3.5-397B-A17B 1,550 397B MoE, 17B active
Qwen/Qwen3.5-35B-A3B 1,486 35B MoE, 3B active

Qwen3-TTS & Specialist Models

Beyond the core language models, the Qwen family extends into specialist domains:

  • Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice — 1,873 likes, text-to-speech with custom voice support
  • Qwen/Qwen3-Coder-Next — 1,585 likes, next-gen coding model
  • Qwen/Qwen3-Coder-480B-A35B-Instruct — 1,359 likes, massive coding MoE

⚙️ Engine Updates

llama.cpp — b10375 (Today!)

The llama.cpp project is on fire with 3 builds today alone on August 12:

Build Time (UTC) Highlights
b10375 12:18 chat: tighten bare function parsing for Qwen models (#26793)
b10373 09:23 imatrix.cpp: move finite check, only check touched experts (#26861)
b10369 04:52 mtmd: support pocket-tts — transposed convolutions as GEMM + col2im (#26871)

What matters today: The b10375 Qwen function parsing improvement directly supports the new Qwen3.8 release. The expert quantization optimization (b10373) benefits all MoE models, including the new Qwen3.8-2.4T. The pocket-tts support (b10369) opens up local TTS inference — potentially complementary to Qwen3-TTS.

RTX 3060 (12GB): Qwen3.8-4B-Distilled at Q8_0 (~5GB) or Qwen3.6-27B at Q4_K_M (tight at ~15GB — needs swap). Stick with the 4B distilled or Qwen3.5-9B for comfortable inference.

RTX 3090 (24GB): Qwen3.8-27B at Q4_K_M (~15GB) with room for context. Qwen3.6-35B-A3B at Q4_K_M (~19GB). Qwen3-Coder-30B-A3B-Instruct at Q6_K (~22GB).

Ollama — 3 Releases This Week (v0.32.7 → v0.32.9)

Ollama had a massive week:

Version Date Headline
v0.32.9 Aug 11 NVIDIA Nemotron 3.5 Lightning — 30B MoE with 3B active params for always-on agents
v0.32.8 Aug 10 Muse Glimmer — full cross-platform support for Meta’s agent-focused model
v0.32.7 Aug 10 Muse Glimmer initial MLX support on Apple Silicon

NVIDIA Nemotron 3.5 Lightning is the standout: an open 30B MoE designed for the execution layer of always-on agents, built for use with OpenClaw and Hermes Agent via the NVIDIA NemoClaw stack.

Muse Glimmer — Meta Superintelligence Labs’ first open model release — is a 30B multimodal model purpose-built for agent workloads that run locally. Ollama’s MLX engine delivers SOTA performance on Apple Silicon.

vLLM — v0.27.0 (August 10)

A massive release with 561 commits from 242 contributors (64 new):

  • Kimi K3 full-stack support: Core model files, Python and Rust frontends, AttnRes kernels, DeepGEMM support, compressed-tensors quantized checkpoints, DSpark AR fusion, and shared expert sharding
  • DSpark Markov heads support in the v0.27.1 patch (August 11)
  • This is the release that makes the 2.8T-parameter Kimi-K3 production-ready for inference

SGLang — v0.5.17 (August 8)

582 PRs from 194 contributors with day-0 Kimi K3 support:

  • Serving the 2.8T-parameter LatentMoE (896 experts, top-16) natively
  • DSpark speculative decoding, chunked-prefill PP with TP decode
  • KDA-aware prefix caching, HiCache L2 over DCP
  • 383.7 tok/s at accept length ~5 on DeepSeek-V4-Pro, TP8 on B300 (from v0.5.16’s DSpark algorithm)

📰 AI News (Hacker News)

Today’s AI-related HN stories:

Score Story Comments
511 AI is removing the middle class of software engineering 430
444 DeepSeek V4 Pro 0813 153
271 Qwen3.8-2.4T-A95B 68
188 Grok 4.6 Scores 61 on AII Index 169

The “AI removing the middle class of software engineering” post dominated discussion (511 points, 430 comments) — a continuation of the ongoing debate about AI’s impact on developer roles.

🖥️ Hardware Recommendations

RTX 3060 (12GB)

  • Sweet spot: Qwen3.8-4B-Distilled (Q8_0), Qwen3.5-9B, Qwen3-TTS-12Hz-1.7B
  • Stretch: Qwen3.6-27B at Q4_K_M (may need CPU offload)
  • Best experience: Run Kokoro-82M for TTS alongside any model

RTX 3090 (24GB)

  • Headline: Qwen3.8-27B at Q4_K_M (~15GB) — the day’s most exciting model at consumer scale
  • Powerhouse: Qwen3.6-35B-A3B, Qwen3.5-397B-A17B (at lower quantization)
  • Agent workloads: Muse Glimmer via Ollama, Nemotron 3.5 Lightning

🔮 What to Watch

  • Qwen3.8 official 27B release: Community forks exist but an official Qwen team release may follow
  • vLLM 0.28: Expected to bring further Kimi K3 optimizations and Qwen3.8 support
  • llama.cpp Qwen3.8 MoE support: The b10375 function parsing update is just the first step
  • Grok 4.6 open-weight: xAI has not released weights yet — benchmark results suggest strong competitive positioning

Sources: HuggingFace API, llama.cpp GitHub, Ollama GitHub, vLLM GitHub, SGLang GitHub, Hacker News API, Artificial Analysis

Scan completed: 2026-08-12T12:30Z