Qwen-Drive 1.0 Lands, DFlash2 Drafts the 27B, and llama.cpp Patches NV Vulkan
Qwen pushes into physical AI with Qwen-Drive 1.0 and Qwen-RobotManip, the community ships DFlash2 speculative-decoding draft models for the 27B workhorse, and llama.cpp lands three builds including an NVIDIA Vulkan queue-submission workaround.
🔥 Top Stories
Qwen leaves the chatbox: Qwen-Drive 1.0 and Qwen-RobotManip land
Two new repos from QwenLM put the model family in the physical world:
- Qwen-Drive-1.0 (380 ⭐) — “An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving.”
- Qwen-RobotManip (166 ⭐) — the official robot-manipulation repo.
Why it matters: every major driving stack in production is closed — Waymo’s end-to-end EMMA models, Tesla’s FSD — and the open driving-VLM work out there is mostly academic. Qwen is the first frontier model family to ship an explicit autonomous-driving VLM foundation model under its own name, and pairing it with a robot-manipulation repo on the same day makes the intent clear: embodied AI, not just chat. It’s also the third leg of this week’s Qwen expansion — the Qwen-Image debut (8,308 ⭐) two days ago, the speech stack (Qwen3-TTS, 13,372 ⭐; Qwen3-ASR, 3,519 ⭐) yesterday, and now the physical world. The Qwen3.8-27B workhorse (#1 on HuggingFace, 14,912 likes) stays the center of gravity, and a new official Qwen3.8 series repo (4,103 ⭐) anchors the family. For self-hosters the near-term value is an inspectable, fine-tunable VLM base for the hardest autonomy domain; the roadmap value is the destination.
The community starts drafting the 27B: DFlash2 speculative-decoding models
Two community repos tagged dflash2 + speculative-decoding now sit on the Qwen tracker for the #1 model on HuggingFace:
- incoai/Qwen3.8-27B-DFlash2 — 231 likes, 364,919 downloads (new today)
- z-lab/Qwen3.8-27B-DFlash2 — 272 likes, 338,620 downloads
What they are: speculative-decoding draft models — a small drafter proposes tokens and the 27B verifies them in a single forward pass, so decode speed goes up without quality loss (verification preserves the target model’s distribution).
Why it matters: two independent groups shipping DFlash2 drafts for the same 27B model in the same day is a momentum signal — the speculative-decoding wave that has been living at the frontier tier is arriving at the consumer 27B tier. It lands on top of the MTP-quantized community GGUFs (HauhauCS Aggressive-MTP, 2.16M downloads; JonathanColetti MTP, 2.99M downloads) and llama.cpp’s daily build cadence — the self-hosting stack for Qwen3.8-27B is being tuned from every direction at once: quants, MTP, and now draft heads. On 24GB-class hardware the target is interactive decode speed on the model that already fits.
llama.cpp ships three builds — b10938 is the NVIDIA Vulkan fix
Three builds landed today (05:14–10:34 UTC):
- b10937 — opencl: apply the noshuffle row-alignment rule to q4_K, q5_K and q8_0, not just q6_K (#28575)
- b10938 — vulkan: workaround NV queuesubmit driver bug (#28830)
- b10941 — tests : reduce FA test sizes (#28842)
Why it matters: b10938 is the user-visible one. NVIDIA’s Vulkan driver has a bug where two queues on the same device submitting simultaneously can break internal synchronization; the workaround is a mutex around queue submission, and it fixes the multibatch (-j) breakage reported in the #28740 discussion. Multibatch is how you get real throughput out of a single card, so this unblocks a core Vulkan use case on NVIDIA hardware. b10937 extends an existing row-alignment correctness rule to three more quantized dtypes on OpenCL — Intel iGPU and AMD-without-ROCm users get properly aligned q4_K/q5_K/q8_0 kernels. b10941 is CI hygiene: the flash-attention tests took tens of minutes on Vulkan + COOPMAT1 runners and were stalling CI. In flight: #28819 (CUDA contiguous-rows unary kernels), #28818 (ELMOD 2.7b — a German research model on a GPTNeoX base), and the XDNA (AMD NPU) backend first spotted yesterday.
📊 Model Trends
HuggingFace trending (top 15 by likes, 2026-09-13):
| Model | Likes | Downloads | Notes |
|---|---|---|---|
| Qwen/Qwen3.8-27B | 14,912 | 7.77M | #1 — the local workhorse |
| black-forest-labs/FLUX.1-dev | 14,638 | 823K | #2 — top text-to-image |
| deepseek-ai/DeepSeek-R1 | 13,627 | 649K | reasoning staple |
| moonshotai/Kimi-K3 | 11,310 | 2.21M | — |
| stabilityai/stable-diffusion-xl-base-1.0 | 8,140 | 2.40M | — |
| CompVis/stable-diffusion-v1-4 | 7,065 | 603K | — |
| meta-llama/Llama-3.1-8B-Instruct | 6,930 | 5.57M | 8B backbone |
| hexgrad/Kokoro-82M | 6,864 | 11.57M | 82M-param TTS |
| meta-llama/Meta-Llama-3-8B | 6,657 | 303K | — |
| openai/whisper-large-v3 | 6,269 | 4.86M | STT workhorse |
| sentence-transformers/all-MiniLM-L6-v2 | 5,907 | 252.9M | embedding glue — most-downloaded on the list |
| black-forest-labs/FLUX.1-schnell | 5,805 | 739K | fast T2I variant |
| deepseek-ai/DeepSeek-V4-Pro | 5,531 | 616K | the V4 line |
| Tongyi-MAI/Z-Image-Turbo | 5,242 | 691K | Alibaba’s Z-Image line |
| MiniMaxAI/MiniMax-H3 | 5,216 | 4.82M | text-to-video — first video model in the top 15 |
Trend notes:
- Qwen3.8-27B still holds #1 at 14,912 (+104 overnight), now 274 ahead of FLUX.1-dev (14,638) — a text model sitting above the top image model, extending yesterday’s 190-point lead to 274.
- The fresh signal is video: MiniMax-H3 at #15 (5,216 likes, 4.82M downloads) is a text-to-video / image-to-video model — the first video-generation model in the top 15 of a text-models list. Video is starting to compete for the top of the list.
- The utility tail is still long: all-MiniLM-L6-v2 at 252.9M downloads, and the 2019 openai-community/gpt2 still climbing at #30 (4,021 likes, +59 overnight, 15.16M downloads) — a legacy checkpoint pulled constantly as a fine-tuning base.
Qwen Tracker
| Model | Likes | Downloads | Status |
|---|---|---|---|
| Qwen3.8-27B (dense) | 14,912 | 7.77M | #1 on HF — the local workhorse |
| Qwen3.8-27B-FP8 | 793 | 7.85M | 8-bit — still out-downloads the base |
| Qwen3.8-Flash-Next | 5,155 | 624K | qwen4_exp foundation model |
| Qwen3.8-2.4T-A95B (MoE flagship) | 1,217 | 53.3K | 2.4T total / 95B active — cluster-class |
| unsloth/Qwen3.8-27B-GGUF | 3,979 | 11.01M | Community quant — most-downloaded overall |
| incoai/Qwen3.8-27B-DFlash2 | 231 | 365K | New — speculative-decoding draft |
| z-lab/Qwen3.8-27B-DFlash2 | 272 | 339K | Speculative-decoding draft |
| ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF | 935 | 770K | Mixed-precision research quant |
| huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF | 697 | 2.68M | Abliterated |
The movement is in the repos and the drafts, not new weights. No new official Qwen 3.8 model IDs today — the energy is the Qwen-Drive-1.0 + Qwen-RobotManip physical-AI push and the first community DFlash2 draft heads for the 27B (see Top Stories). The 8-bit trend keeps going: Qwen3.8-27B-FP8 (7.85M) still out-downloads the base BF16 (7.77M). The uncensored/abliterated tier stays deep — orcarouter’s Uncensored-FP8 (1,568) leads it, with MTP-enabled quants doing the heavy download lifting.
⚙️ Engine Updates
- llama.cpp — see Top Stories (b10937, b10938, b10941). The user-visible wins: b10938’s NVIDIA Vulkan queue-submission workaround (fixes
-jmultibatch) and b10937’s OpenCL row-alignment fix for q4_K/q5_K/q8_0. - Ollama — v0.34.0 (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)
- Everyone should slow down AI development except for me — 586 pts, 347 comments (discussion, source). The day’s top story — a personal essay on the pace of AI development where the title is the argument: everyone should slow down, except the author.
- Nvidia is the central bank of AI — 510 pts, 358 comments (discussion, source). The Economist’s interactive briefing frames Nvidia as the institution that sets compute “liquidity” for the whole AI economy — a useful counterweight to the open-model momentum dominating this site’s own pages.
- Why are AI agents lying, cheating and coordinating? — 335 pts, 391 comments (discussion, source). Yoshua Bengio’s new publication on multi-agent misalignment — the comment count beating the score says it’s being argued hard, and it lands the day after the RubyGems agent-attack story topped the front page. Agent safety is the week’s recurring theme.
- Model fatigue sets in — CNBC on the “dizzying” release cadence across Anthropic, OpenAI, Meta and Google this week. Context, not a single event.
🔄 What Changed Since Last Scan
Diff vs 2026-09-12 (12 new pipeline items: 3 llama.cpp releases + 3 HN front-page stories + 1 new Qwen model + 5 new Qwen repos; 0 new HF trending entries):
- 3 llama.cpp builds (all Sep 13): b10937 (OpenCL row-alignment for q4_K/q5_K/q8_0, #28575), b10938 (NVIDIA Vulkan queue-submission workaround, #28830), b10941 (FA test-size reduction, #28842).
- 5 new Qwen repos: Qwen-Drive-1.0 (380 ⭐), Qwen-RobotManip (166 ⭐), the official Qwen3.8 series repo (4,103 ⭐), Qwen3.8-Flash-Next (358 ⭐), and Qwen-VL (6,725 ⭐ — the original 2023 vision-language repo re-trending, not a new release).
- 1 new Qwen model: incoai/Qwen3.8-27B-DFlash2 (231 likes, 365K downloads) — a speculative-decoding draft head for the 27B workhorse, alongside z-lab’s DFlash2 twin (272 likes, 339K).
- 3 new HN front-page stories: the slowdown essay (586), the Nvidia-central-bank briefing (510), and Bengio’s agent-misalignment publication (335).
- No new HF trending entries. Qwen3.8-27B holds #1 at 14,912 (+104); star drift across the Qwen repos: Qwen3-TTS 13,360→13,372, qwen-code 27,797→27,814, Qwen-Image 8,302→8,308, Qwen3-ASR 3,517→3,519. Today’s movement: Qwen’s physical-AI push, the first DFlash2 draft heads for the 27B, and llama.cpp’s NVIDIA Vulkan fix.
Sources: HuggingFace API · llama.cpp releases · vLLM releases · Ollama releases · SGLang releases · QwenLM GitHub · Hacker News Research completed: 2026-09-13T13:04Z