Nvidia Buys Hugging Face for $12.9B, Mistral Banks €3B, Vulkan Gets Faster
Nvidia and Hugging Face announce a $12.9B acquisition, Mistral closes a €3B Series D at a €21B+ valuation led by Samsung, and llama.cpp lands a wave of Vulkan kernel-fusion performance builds.
🔥 Top Stories
Nvidia Buys Hugging Face for $12.9B
The week’s biggest deal closed as an announcement: per CNBC, Nvidia “officially agreed to buy open-source AI platform Hugging Face for $12.9 billion.”
Why it matters: Hugging Face is the distribution layer for everything in today’s briefing — every model in the trending table below lives on HF, and the HF API powers the tracker this post is built from. Nvidia now owns both the compute and the registry. The strategic direction of HF’s tooling is the variable to watch: TGI (HF’s own inference server) has already been put into maintenance mode since December 2025 — a signal of where HF’s engineering attention is going.
Mistral Banks €3B — Samsung Leads
Mistral closed a €3B Series D at a post-money valuation above €21 billion, led by Samsung Electronics with Scaleup Europe Fund (EQT) and existing investor PSG Equity as co-leads. New money came from Advent, funds and accounts managed by BlackRock, and the Grand Duchy of Luxembourg; existing participants include NVIDIA, a16z, ASML, DST Global, and Lightspeed. The HN thread (610 points, 429 comments) is the most-discussed story of the day.
Why it matters: watch the pattern — hardware makers are buying into the open-model layer: Samsung into Mistral, Nvidia into Hugging Face. Proceeds go to “frontier research” and “compute capacity for training” per the release — i.e., more open-weight frontier models trained on European soil, which is the “sovereign AI” pitch in one sentence.
llama.cpp’s Vulkan Week
Three builds in a day: b10857, b10858, b10859.
- b10858 is the headline: Vulkan UNARY(GELU|SIGMOID|SILU|SOFTPLUS)+MUL fusion (#27220) — fusing activation+mul pairs into a single shader pass cuts memory round-trips on the Vulkan backend (AMD, Intel, iGPU, mobile GPUs). A real performance win for non-CUDA self-hosters.
- b10857: fixes Vulkan-Hpp handle usage on 32-bit targets (#22892).
- b10859: clang/musl header fixes (#28566) — unblocks musl-based distros.
- In flight: PR #28593 fixes three null-pointer crashes on pre-1.2 Vulkan drivers that report 1.2 features (reproduced on a Galaxy S21 FE / Mali-G78 / Android 16) — open, not yet merged. PR #28585 adds Hexagon RELU/LEAKY_RELU ops (merged) — more ops for the Qualcomm NPU backend.
- Also spotted in the PR stream: openbmb/MiniCPM5-2B just went public, and the llama.cpp community is already wiring up its chat template (#28595).
“Google Jail” for Brand-New Domains
Per weirdgloop’s write-up (HN, 262 points): Google is effectively suppressing brand-new domains — the author reports ~90% of wikis that launched on new domains since the March 2024 core update don’t show up in search beyond the main page, sometimes for up to a year. Workaround: publish under a subdomain of an established domain. Not strictly an AI story, but it’s the “SEO slop” cleanup reshaping the index — relevant to anyone self-hosting docs or content that needs to be found.
Model Fatigue Week
Per CNBC, all four frontier labs shipped this week: Anthropic (Claude Fable 5.1, Claude Mythos 5.1), Meta (Muse Spark 1.3), Google (Gemini 3.8 Flash), OpenAI (GPT-6 Astra). The llm-stats tracker lists all of them. The churn is the story: frontier APIs move weekly while the local side consolidates around one 27B workhorse — Qwen3.8-27B has pulled 6.7M downloads.
📊 Model Trends
HuggingFace trending (top 10 by likes, 2026-09-08):
| Model | Likes | Downloads | Notes |
|---|---|---|---|
| black-forest-labs/FLUX.1-dev | 14,518 | 790K | T2I king holds #1 |
| Qwen/Qwen3.8-27B | 14,336 | 6.71M | #2 — 182 likes behind #1 |
| deepseek-ai/DeepSeek-R1 | 13,618 | 777K | — |
| moonshotai/Kimi-K3 | 11,236 | 2.38M | — |
| stabilityai/stable-diffusion-xl-base-1.0 | 8,123 | 1.95M | — |
| hexgrad/Kokoro-82M | 6,832 | 11.5M | 82M-param TTS; most-downloaded model on the list |
| openai/gpt-oss-120b | 5,171 | 5.41M | MXFP4, ~61GB — multi-GPU home-lab reference |
| zai-org/GLM-5.2 | 5,077 | 965K | MoE with glm_moe_dsa sparse-attention tag |
| MiniMaxAI/MiniMax-H3 | 5,031 | 5.0M | Text/image-to-video |
| Qwen/Qwen3.8-Flash-Next | 4,999 | 503K | Tagged qwen4_exp |
Trend notes:
- The 27B sweet spot, and the quant is the artifact: Qwen3.8-27B’s official repo (6.7M dl) is actually shadowed by unsloth’s GGUF at 10.7M downloads. Q4_K_M ≈ 15–16GB: fits a 24GB card (RTX 3090/4090) with headroom; 12GB cards need partial offload.
- TTS is quietly huge: Kokoro-82M has 11.5M downloads for an 82M-parameter model — it runs on anything, so people just download it.
- The gpt-oss split: gpt-oss-120b (5.4M dl) for multi-GPU labs; gpt-oss-20b (5,007 likes, 6.55M dl) is the single-GPU variant.
- DSA is spreading: GLM-5.2’s
glm_moe_dsatag (DeepSeek Sparse Attention) — the architecture DeepSeek introduced is now showing up in other labs’ MoE releases.
Qwen Tracker
| Model | Likes | Downloads | Status |
|---|---|---|---|
| Qwen3.8-27B (dense) | 14,336 | 6.71M | The local workhorse |
| Qwen3.8-2.4T-A95B (MoE flagship) | 1,205 | 52.7K | 2.4T total / 95B active — cluster-class, FP8 ≈ 2.4TB |
| Qwen3.8-Flash-Next | 4,999 | 503K | qwen4_exp tag; FP8 variant at 249K dl |
| unsloth/Qwen3.8-27B-GGUF | 3,667 | 10.7M | Community quant; NVFP4 at 3.26M dl |
Community ecosystem is in full swing: abliterated/uncensored variants (orcarouter, OBLITERATUS, huihui-ai, JonathanColetti), z-lab’s DFlash2 speculative-decoding variant, ISTA’s GSQ-RCO mixed-precision GGUF. On the tooling side: qwen-code (27.7K ⭐) and the new E-CommerceBench — a long-horizon benchmark where 18 LLM agents each run a simulated online store for 365 days on real market data. No new Qwen 3.8 model IDs since yesterday.
⚙️ Engine Updates
- llama.cpp — see Top Stories (b10857–b10859). In flight: OpenCL prefill512 FA compile fix (#28605).
- Ollama — v0.34.0-rc1 (Sep 5) is still RC. The headline feature — using Ollama models in ChatGPT Desktop on macOS — shipped in RC form alongside Apple Silicon structured-output performance improvements and image support via compacted responses. GA is the event to watch.
- SGLang v0.5.19 — released Sep 5: 786 PRs from 214 contributors; adds Qwen3.8 (2.4T-A95B) and Qwen3.8-27B to the cookbook — day-0 serving for Qwen’s MoE flagship.
- vLLM v0.28.0 — Aug 26: 584 commits; a Kimi-K3 performance push (DCP support, fused FlashKDA decode/prefill kernels, 1.5–3× all-gather speedups).
- TGI — v3.3.7 remains in maintenance mode — the backdrop to today’s Nvidia/HF story.
📰 AI News (Hacker News)
- Mistral €3B — 610 pts, 429 comments (discussion).
- Google Jail — 262 pts, 95 comments (discussion).
- TradingAgents — a multi-agent LLM financial trading framework hit the front page (81 pts). Agent benchmarks are moving toward long-horizon economic tasks — cf. QwenLM’s E-CommerceBench above.
- Arm Mali G2-Ultra NX — Arm pitches its AI-native mobile GPU for desktop-class mobile gaming (59 pts). More NPU-class silicon in phones = more on-device targets for llama.cpp (see today’s Hexagon ops).
- Applied Compute — an end-to-end open-weight training + inference stack (GRPO/SFT, inference routing, evals) is quietly gaining traction (26 pts).
🔄 What Changed Since Last Scan
Diff vs 2026-09-07 (9 new items):
- 3 llama.cpp builds: b10857 (32-bit Vulkan handle fix #22892), b10858 (Vulkan UNARY+MUL fusion #27220), b10859 (clang/musl header fix #28566).
- 6 new HN front-page stories: Mistral €3B (610), Google Jail (262), TradingAgents (81), reCAPTCHA/Firefox (77), Arm Mali G2-Ultra NX (59), Applied Compute (26).
- No new HF trending model IDs; no new Qwen 3.8 models — the Qwen tracker is steady: 27B dense, 2.4T-A95B MoE, and Flash-Next (qwen4_exp) are all unchanged since yesterday.
Sources: HuggingFace API · llama.cpp releases · Hacker News · Mistral · CNBC Research completed: 2026-09-08T13:02Z