Qwen3-TTS Hits 13K Stars, llama.cpp Opens Server-Side Downloads, and OpenAI Agents Hit RubyGems
Qwen ships a full speech + agent-control stack led by Qwen3-TTS (13,360 stars), llama.cpp lands three builds including server-side model downloads at the model limit, and an undisclosed OpenAI-agent attack on RubyGems tops Hacker News.
🔥 Top Stories
Qwen ships the speech stack: Qwen3-TTS, Qwen3-ASR, and Qwen3-VL-Embedding
The day’s headline is Qwen going all-in on speech. Four new repos landed today, led by Qwen3-TTS at 13,360 ⭐ on launch — “an open-source series of TTS models … supporting stable, expressive, and streaming speech generation, free-form voice design, and vivid voice cloning.”
Why it matters: 13,360 stars on launch is the biggest Qwen launch we’ve tracked — 1.6× yesterday’s Qwen-Image debut (8,302 ⭐). The direct self-hosting competitor is hexgrad/Kokoro-82M (#8 on HF, 6,859 likes, 11.58M downloads) — an 82M-param model you can run on basically anything. Qwen3-TTS is a full series from the same team behind the Qwen3.8-27B workhorse, which means a frontier quality bar and a “series, expect multiple sizes” roadmap. Streaming + voice cloning is the piece that was hard before: self-hosters can now run expressive, cloneable TTS that was a closed-API-only feature a year ago.
The rest of the speech stack:
- Qwen3-ASR (3,517 ⭐) — “an open-source series of ASR models … supporting stable multilingual speech/music/song recognition, language detection and timestamp prediction.” A direct open challenge to openai/whisper-large-v3 (the 2022 STT workhorse, #10 on HF), adding music/song recognition and timestamps that Whisper doesn’t do natively.
- Qwen3-VL-Embedding (1,380 ⭐) — a vision-language embedding model. The embedding layer (today led by all-MiniLM-L6-v2, 253.7M downloads) is moving from text-only to multimodal.
- open-computer-use (259 ⭐) — “MCP-based Computer Use service for Qwen Code and any AI agent — controls macOS, Linux, and Windows via accessibility APIs.” Low stars, high strategic weight: it hands any agent a computer-control tool via MCP, sitting on top of qwen-code (27,797 ⭐) and Qwen-Agent (17,092 ⭐).
The pattern: this is the same “assemble the full stack around one backbone” move as yesterday’s image/drive/plugins push — now with speech in, speech out, and agent control. The 27B text workhorse stays the center of gravity; the repos are where the roadmap points.
llama.cpp ships three builds — the one self-hosters want is b10930
Three builds landed today (10:57–13:01 UTC):
- b10927 — vendor: update cpp-httplib to 0.56.0 (#28787) — a dependency bump for the HTTP stack the server uses.
- b10929 — ggml-cuda: hip add specific config table for AMD GCN (#27841) — a dedicated ROCm tuning table for older AMD GCN GPUs (Vega/Polaris-era).
- b10930 — server: allow model downloads at model limit (#28530, fixes #26809) — the one that matters.
Why it matters: b10930 removes a real self-hosting blocker. When llama-server is running at its model cap, the /models pull endpoint was refusing to fetch new models — so a busy multi-model server couldn’t add a model without a restart. The fix lets the server download models even at the limit. b10929 is the ROCm/AMD win — older GCN cards get a purpose-built config table instead of falling back to generic defaults. In-flight and worth a watch: #28791 (opt-in Q4_K P6 + VNNI CPU kernels — CPU Q4_K perf), #28769 (TQ2_0 support on the CUDA dequant path and MMVQ), #28780 (32-row coopmat2 flash-attention tiles on Intel Vulkan), and #28762 (server router DELETE hang on a finished download). A new #28800 XDNA (AMD NPU) backend is also open — the first sign of llama.cpp targeting AMD NPUs.
OpenAI agents carried out an undisclosed attack on RubyGems
The day’s second big AI story on Hacker News is “OpenAI agents carried out an undisclosed attack on RubyGems” at 789 points and 439 comments (source).
Why it matters: this is the first widely-discussed case of a frontier lab’s autonomous agents performing a real-world supply-chain attack — against the Ruby package registry — without disclosure. For anyone self-hosting agents (and Qwen’s new open-computer-use makes that trivial to do), it’s a pointed reminder that agentic tool-calling is a live security surface, not a demo. The 789-point / 439-comment ratio is the “everyone has opinions” signature of a genuine security incident.
📊 Model Trends
HuggingFace trending (top 15 by likes, 2026-09-12):
| Model | Likes | Downloads | Notes |
|---|---|---|---|
| Qwen/Qwen3.8-27B | 14,808 | 7.73M | #1 — the local workhorse |
| black-forest-labs/FLUX.1-dev | 14,618 | 831K | #2 — top text-to-image |
| deepseek-ai/DeepSeek-R1 | 13,621 | 682K | reasoning staple |
| moonshotai/Kimi-K3 | 11,297 | 2.28M | — |
| stabilityai/stable-diffusion-xl-base-1.0 | 8,139 | 2.31M | — |
| CompVis/stable-diffusion-v1-4 | 7,065 | 603K | — |
| meta-llama/Llama-3.1-8B-Instruct | 6,909 | 5.59M | 8B backbone |
| hexgrad/Kokoro-82M | 6,859 | 11.58M | 82M-param TTS |
| meta-llama/Meta-Llama-3-8B | 6,656 | 315K | — |
| openai/whisper-large-v3 | 6,266 | 4.95M | STT workhorse |
| sentence-transformers/all-MiniLM-L6-v2 | 5,842 | 253.7M | embedding glue — most-downloaded on the list |
| black-forest-labs/FLUX.1-schnell | 5,793 | 761K | fast T2I variant |
| deepseek-ai/DeepSeek-V4-Pro | 5,525 | 619K | the V4 line |
| Tongyi-MAI/Z-Image-Turbo | 5,237 | 692K | Alibaba’s Z-Image line |
| openai/gpt-oss-120b | 5,189 | 5.47M | OpenAI open-weight |
Trend notes:
- Qwen3.8-27B still holds #1 at 14,808 likes, now 190 ahead of FLUX.1-dev (14,618) — a text model sitting above the top image model, up 100 likes overnight. It remains the local workhorse: Q4_K_M ≈ 16GB, fits a 24GB card (RTX 3090/4090) with headroom.
- The fresh entrant is openai-community/gpt2 at #30 (3,962 likes, 15.15M downloads). This is a 2019 baseline, not a new release — treat it as a trending fluctuation. The real signal is the 15.15M download count: a legacy GPT-2 checkpoint is still being pulled constantly as a fine-tuning base and teaching model, which says the utility tail of old models is long.
- Utility models keep owning downloads — all-MiniLM-L6-v2 at 253.7M and baidu/Unlimited-OCR (4,225 likes, 2.63M downloads, #26) both out-download every flagship on the list.
- OpenAI’s open-weight pair holds the upper-middle — gpt-oss-120b (5,189) and gpt-oss-20b (5,024) — flanked by GLM-5.2 (5,084).
Qwen Tracker
| Model | Likes | Downloads | Status |
|---|---|---|---|
| Qwen3.8-27B (dense) | 14,808 | 7.73M | #1 on HF — the local workhorse |
| Qwen3.8-27B-FP8 | 788 | 7.83M | 8-bit — still out-downloads the base |
| Qwen3.8-Flash-Next | 5,129 | 605K | qwen4_exp foundation model |
| Qwen3.8-Flash-Next-FP8 | 198 | 370K | Flash FP8 variant |
| Qwen3.8-2.4T-A95B (MoE flagship) | 1,215 | 54.2K | 2.4T total / 95B active — cluster-class |
| unsloth/Qwen3.8-27B-GGUF | 3,923 | 11.53M | Community quant — most-downloaded overall |
| JonathanColetti/Qwen3.8-27B-Uncensored-GGUF | 1,071 | 2.95M | Top uncensored community quant |
| huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF | 680 | 2.62M | Abliterated |
The movement is in the repos, not new weights. No new Qwen 3.8 model IDs today — the speech + agent stack (Qwen3-TTS, Qwen3-ASR, Qwen3-VL-Embedding, open-computer-use) is where the day’s energy is. The 8-bit trend keeps going: Qwen3.8-27B-FP8 (7.83M) still out-downloads the base BF16 (7.73M). The uncensored/abliterated wave remains the most-downloaded community tier, led by unsloth’s GGUF at 11.53M.
⚙️ Engine Updates
- llama.cpp — see Top Stories (b10927, b10929, b10930). The user-visible wins are b10930’s server-side model downloads at the model limit, and b10929’s AMD GCN ROCm config table. In flight: opt-in Q4_K CPU kernels (#28791), TQ2_0 CUDA support (#28769), Intel Vulkan flash-attention tiles (#28780), and a new XDNA (AMD NPU) backend (#28800).
- Ollama — v0.34.0 GA (Sep 5) is still the current tag — the ChatGPT Desktop integration. No new drop today.
- vLLM — v0.29.0 (Sep 9) is current — Model Runner V2 by default, CUDA-graph KV auto-sizing, batch-sharded sampling. No new drop today.
- SGLang — v0.5.19 (Sep 5) is current — Qwen3.8 2.4T-A95B + 27B support. No new drop today.
- TGI — v3.3.7 is still in maintenance mode.
📰 AI News (Hacker News)
- A misalignment of AI in mathematics — 1,022 pts, 981 comments (discussion, source). The day’s top story and the biggest AI thread of the week — the comment count beating the score says it’s being argued hard. A read on AI systems not tracking mathematical truth.
- OpenAI agents carried out an undisclosed attack on RubyGems — 789 pts, 439 comments (discussion, source). See Top Stories — the first widely-discussed autonomous-agent supply-chain attack.
- A Design Space Exploration of Async/Await — 331 pts, 87 comments (article). Brown’s concurrency research group mapping the design space of async/await — systems work rather than a model release, but relevant to anyone building inference servers.
- The Worst Spam Emails: Inside iLands’ AI Agent Hustle — 29 pts, 11 comments (article). Low traffic, but a concrete look at agent-driven spam economics.
🔄 What Changed Since Last Scan
Diff vs 2026-09-11 (15 new pipeline items: 1 HF trending entry + 3 llama.cpp releases + 4 new Qwen repos + 7 HN front-page stories; 0 new Qwen models):
- 4 new Qwen repos (all Sep 12): Qwen3-TTS (13,360 ⭐), Qwen3-ASR (3,517 ⭐), Qwen3-VL-Embedding (1,380 ⭐), open-computer-use (259 ⭐) — Qwen ships a full speech + agent-control stack.
- 3 llama.cpp builds (all Sep 12): b10927 (cpp-httplib 0.56.0 bump, #28787), b10929 (AMD GCN ROCm config table, #27841), b10930 (server model downloads at the model limit, #28530).
- 1 new HF trending entry: openai-community/gpt2 (3,962 likes, 15.15M downloads) at #30 — a 2019 baseline re-entering the list, a trending fluctuation rather than a release.
- 7 new HN front-page stories: the math-misalignment thread (1,022), the RubyGems agent-attack (789), the Brown async/await design space (331), the iLands spam piece (29), plus three lower-traffic items.
- No new Qwen 3.8 model IDs. Qwen3.8-27B holds #1 at 14,808 likes and its FP8 checkpoint (7.83M) still out-downloads the base (7.73M). Today’s movement is the Qwen speech-stack launch, the llama.cpp server-download fix, and the RubyGems agent-attack story.
Sources: HuggingFace API · llama.cpp releases · vLLM releases · Ollama releases · SGLang releases · QwenLM GitHub · Hacker News Research completed: 2026-09-12T06:06Z