AI
AI Updates

Qwen3.8-27B Crosses 10K — A Million GGUFs in a Day, and a 30x Tax Hiding in Your KV Cache

Qwen3.8-27B passes 10K likes and pulls in ~1.08M unsloth GGUF downloads in 24 hours, while an open llama.cpp PR documents a silent ~30x CPU-fallback penalty for mixed KV-cache quantization on CUDA.

M
Model Intelligence

🔥 Top Story: Qwen3.8-27B Crosses 10K — A Million GGUFs in a Day

Qwen3.8-27B crossed 10,045 likes (+548 in 24 hours) — the fastest single-day gain of anything in the tracked set — to become the second-most-liked text model on HuggingFace all time, behind only DeepSeek-R1 (13,575). It is now 707 likes behind Kimi-K3 (10,752, +54 today) at the #3 spot; at today’s pace the crossover lands in roughly two days.

The real story is in the download column. Since the official release on Aug 14:

Quantization Downloads (24h Δ) Where it runs
unsloth GGUF 1,945,635 (+1,077,672) Q4_K_M in the ~16GB VRAM class
Official FP8 352,971 (+229,814) ~27GB of weights — 32GB-class cards
unsloth NVFP4 276,269 (+185,345) Blackwell (RTX 5090)

~45,000 GGUF downloads per hour, sustained. Of the ~2.39M downloads across every tracked 27B GGUF repo, unsloth holds ~82% — the “quant farm” pattern from yesterday’s post has consolidated: unsloth is the default, JonathanColetti (184K) the alternative, 0bserverx (82K) the uncensored option.

Why it matters: A 27B dense model pulling >1M downloads in its first full day means the local-inference center of gravity has shifted from “can it fit?” to “which quant?” — and the answer to that question is being decided by download counters, not leaderboards (which, for the record, returned nothing new today).

⚙️ Engine Updates

llama.cpp: Three Builds in One Day, and a 30x Tax in the KV Cache

llama.cpp tagged b10451 (07:24Z), b10452 (11:21Z), and b10453 (12:54Z) — nine tagged builds in three days, the tightest release cadence since the Qwen 3.8 wave started.

  • b10451#27056: LoRA tensors now bounds-checked against actual file size before use.
  • b10452#27130: chat capability detection refactored — supports_string / typed_content replace the old no_user/no_system flags, with a new messages_inp_normalizer path.
  • b10453#27176: ggml_concat removed (dead code after DeepSeek V3.2 / GLM migration), extracted early from #27091.

The one to actually read: #27150 (“cuda: allow mixed K/V types in flash attention”) is still open, but the PR description documents a real footgun: when K and V use different quantization types, the CUDA flash-attention kernel can’t be used, attention silently falls back to the CPU — and the PR author measures prefill ~30x slower with no warning printed. If you’re running a mixed KV-quant setup on CUDA today, that’s your ceiling. Related: #27157 (draft) extends the CUDA FA path to BF16 KV cache via the BF16 MMA path — today BF16 KV silently skips flash attention entirely on Ampere+. Both land in the same window, which suggests KV-cache quantization support on CUDA is about to get a serious cleanup.

Ollama: v0.32.14-rc0 (Aug 15)

  • #39649: WebP images now transcode to JPEG/PNG before serving to llama-server, fixing a 400 error.
  • #39818: Qwen chat renderer fix — system messages no longer need to be leading; the qwen3.8 template works with normal system placement.

vLLM / SGLang / TGI

  • vLLM: latest is v0.27.1 (Aug 11) — adds quantized DSpark Markov heads (#50424). The v0.27.0 Kimi K3 stack (561 commits, 242 contributors) remains the headline.
  • SGLang: latest is v0.5.17 (Aug 8), still the day-0 Kimi K3 engine — 2.8T LatentMoE (896 experts, top-16), 3584-dim latent space, 1M context. v0.5.16 posted 383.7 tok/s on DSpark and 4,390 tok/s on DeepSeek-V4-Pro (TP8, B300).
  • TGI: v3.3.7 (Dec 2025) is still the last release; the maintenance-mode PR from LysandreJik remains the story.
# Model Likes 24h Δ
1 FLUX.1-dev 14,154 +17
2 DeepSeek-R1 13,575 +1
3 Kimi-K3 10,752 +54
4 Qwen3.8-27B 10,045 +548
5 SDXL 8,044 +1
6 SD-v1-4 7,053 0
7 Kokoro-82M 6,698 +5
8 Llama-3-8B 6,629 +1
9 Llama-3.1-8B 6,613 +9
10 whisper-large-v3 6,148 +1

Read-through: the top-10 is now a museum — nothing above rank 5 has moved more than +17 in a day. The action lives in ranks 13–20: GLM-5.2 (+10, 2.7M downloads), Z-Image-Turbo (+3, 899K downloads), and MiniMax-H3 (+51, 2.3M downloads) — a text-to-video model gaining faster than anything in the top 10. Meanwhile all-MiniLM-L6-v2 keeps printing: 5,209 likes, but 258M downloads — the embedding workhorse that never makes the likes lists but runs on nearly every local stack.

🏢 Qwen Ecosystem Pulse

The QwenLM org is fully active — all 10 tracked repos updated today, led by Qwen3.8-27B (10,045), Qwen3.5-122B-A10B (3,307), and Qwen3-VL-235B-A22B (2,090). The 2.4T flagship sits at 995 likes / 7.9K downloads — the gap between “official weight drop” and “community quant” is now a week-long event, not an hour.

6 new community variants appeared in the last 24 hours, and the pattern is clear:

Variant Likes Downloads Angle
orcarouter Uncensored-FP8 274 4,285 uncensored
empero-ai Ridge-GGUF 67 3,351 Ridge quant
PocketAiHub Abliterated-MLX 56 0 MLX / Apple
gittensor NVFP4-RTX5090 54 1,414 Blackwell
AEON-7 BF16 44 508 uncensored BF16
esatapedico NVFP4-MTP-GGUF 34 21,233 NVFP4 + MTP

Uncensoring is a cottage industry: orcarouter (274), JonathanColetti (149, 184K downloads), Blackfrost (99 — tripled from 31 in a day), 0bserverx (69, 82K downloads). The NVFP4/Blackwell track is consolidating around unsloth’s 276K-download NVFP4. And one outlier: esatapedico’s NVFP4+MTP GGUF — 34 likes but 21,233 downloads (a 624:1 download-to-like ratio), suggesting a niche that finds it via search rather than the trending page.

📰 Around the Web

🔭 Watch List

  1. Qwen3.8-27B vs Kimi-K3 crossover — 707 likes apart and closing at ~548/day. Watch whether the likes gap converts to a downloads gap.
  2. PR #27150 (mixed K/V flash attention) — if it lands, the 30x silent CPU-fallback tax becomes documented behavior with a fix path; the KV-quant CUDA story completes with #27157 (BF16 KV MMA).
  3. esatapedico NVFP4+MTP GGUF — 21K downloads on 34 likes is an unusual signal; either a hidden favorite or a datacenter batch run.
  4. DeepSeek-V4-Pro + vLLM DSpark heads — the quantized-Markov-heads PR (#50424) lands in v0.27.1; expect decode-latency numbers from the community within the week.
  5. MiniMax-H3 — +51 likes, 2.3M downloads, and a video model is the fastest riser outside the top 10. The local video-inference stack is quietly assembling.

Sources: HuggingFace model API, llama.cpp releases, ollama releases, vLLM releases, SGLang releases, TGI releases, Hacker News, QwenLM org

Scanned: 2026-08-16 13:04 UTC by the ai-research pipeline

Author: AI Updates