llama.cpp Vulkan Sparse-FA for DSV4/GLM + OpenVINO MoE; Ollama 0.34.1 RC; OpenAI Buys Glass Imaging

🔥 Top Stories

1. llama.cpp ships three builds in an hour: Vulkan sparse Flash Attention for DSV4 / GLM

ggml-org/llama.cpp cut three releases between 11:32 and 12:33 UTC on Sep 15, and the headline one is a Vulkan sparse Flash Attention path aimed squarely at the sparse-attention MoE family:

• b10982 — vulkan: support sparse Flash Attention (#28105). Adds sparse Flash Attention support for DSV4/GLM, tunes the implementation, adds tests, avoids a nondeterministic atomicAdd, and adds cm2 decode-vector support (including an f16vec4 binding for the decode vector).
• b10981 — OpenVINO: optimize stateful decode and GPU MoE inference (#28638). Fixes stateful decode for Gemma-4 per-layer-type head sizes, classifies sliding-window layers structurally on interleaved-SWA models, and adds two knobs: GGML_OPENVINO_REQUANT_KQUANT (pick a 4-bit requant target) and GGML_OPENVINO_SPILL_DIR (spill weights to disk).
• b10980 — opencl: add generic ssm_scan (#28881). A generic ssm_scan for OpenCL, which opens the door to state-space (Mamba-style) models on AMD/Intel iGPUs.

Why it matters. Sparse Flash Attention on Vulkan is the piece that makes DeepSeek V4 and GLM-class sparse-attention MoEs practical on non-NVIDIA silicon (AMD Radeon, Apple Silicon, and other Vulkan backends). Paired with the OpenVINO MoE / stateful-decode work, the day is a clear signal that llama.cpp is closing the gap for large MoE and SSM models on consumer and edge hardware.

2. Ollama cuts v0.34.1-rc2 — an MLX-hardening release

Ollama’s v0.34.1-rc2 (landed the evening of Sep 14) is the first release candidate toward the post-v0.34.0 maintenance line, and it is MLX-focused:

• mlxrunner: evicts prefix-cache snapshots from the active conversation, and now checks system free memory and waits for evicted runners before loading the next MLX model.
• mlx: scopes array lifetimes instead of pinning and sweeping.
• llm: keeps the gemma3n projector off the CPU, and raises the token-repeat limit to 100 (returning an error instead of an incomplete result).
• app: fixes the ChatGPT model-selector spacing (the feature v0.34.0 shipped) and refreshes the Apps layout + command copy.

Why it matters. v0.34.0 (Sep 5) was the big ChatGPT-Desktop integration release. This RC tightens the Apple-Silicon / MLX path — the memory-management fixes (prefix-cache eviction, free-memory gating, evicted-runner waits) are exactly the class of bug that bites multi-model MLX users on Macs with tight unified memory. The ChatGPT-selector spacing fix also shows the Desktop integration is still in active polish. Note: this is an RC, not a stable.

3. TechCrunch: OpenAI to buy smartphone camera maker Glass Imaging for $300M

A TechCrunch report (picked up on HN) says OpenAI is acquiring Glass Imaging, a smartphone camera maker, for ~$300M.

Why it matters. If it closes, it is another concrete step toward the long-rumored OpenAI consumer device: buying a camera/optics specialist is a very different move from a model or data-center purchase, and it points to hardware integration (imaging, sensors, on-device capture) rather than another API. Early story — treat the specifics as reported, not confirmed.

4. Qwen3.8-27B’s community GGUF climbs into the HF trending top-30

unsloth/Qwen3.8-27B-GGUF crossed into the HF trending list today at 4,128 likes / 9.46M downloads — the dense 27B community quant is now a top-tier trending item alongside the official checkpoints.

Why it matters. The 27B dense model keeps its grip as the community’s favorite local workhorse, and the fact that a community GGUF (not an official Qwen upload) is trending at this level signals sustained local-deployment demand.

📊 What Changed Since Last Report (Sep 13)

• llama.cpp — three new builds today: b10982 (Vulkan sparse Flash Attention for DSV4/GLM), b10981 (OpenVINO stateful decode + GPU MoE, Gemma-4 fix, 4-bit requant + weight-spill knobs), b10980 (OpenCL generic ssm_scan).
• Ollama — v0.34.1-rc2 (first RC of the 0.34.1 line): MLX prefix-cache eviction, free-memory gating before model loads, gemma3n projector kept off CPU, token-repeat limit → 100, ChatGPT selector spacing fix.
• OpenAI — reported to be acquiring camera maker Glass Imaging for $300M (TechCrunch).
• Hugging Face — unsloth/Qwen3.8-27B-GGUF enters the trending top-30; Qwen3.8-27B still #1 overall (15,204 likes / 7.70M downloads); DeepSeek-V4-Pro keeps climbing (5,558 likes); Kimi-K3 at 11,360 likes.
• Qwen org — new repo FlashQLA (linear-attention kernel library on TileLang, 702 stars); qwen-code at 27,871 stars.

📡 Qwen 3.8+ Tracker

Hugging Face Models

Model	Likes	Downloads	Notes
Qwen/Qwen3.8-27B	15,204	7.70M	Still #1 on HF
Qwen/Qwen3.8-Flash-Next	5,244	668K	qwen4_exp foundation model
unsloth/Qwen3.8-27B-GGUF	4,128	9.46M	New in trending top-30
Qwen/Qwen3.8-2.4T-A95B	1,225	52K	2.4T MoE flagship
Qwen/Qwen3.8-27B-FP8	827	7.79M	FP8 of the 27B
orcarouter/Qwen3.8-27B-Uncensored-FP8	1,613	464K	Abliterated FP8
OBLITERATUS/Qwen3.8-27B-OBLITERATED	1,219	1.27M	Abliterated
HauhauCS Aggressive-MTP GGUF	1,190	2.30M	Aggressive-MTP GGUF
incoai/Qwen3.8-27B-DFlash2	234	386K	Speculative-decode draft head (up since Sep 13)

Qwen GitHub Repos

Repo	Stars	Updated	Notes
QwenLM/qwen-code	27,871	Sep 15	Terminal coding agent
QwenLM/Qwen	21,735	Sep 15	Main repo
QwenLM/Qwen-Agent	17,099	Sep 15	Agent framework (MCP, RAG)
QwenLM/Qwen3-Coder	16,840	Sep 15	Code model
QwenLM/Qwen-VL	6,727	Sep 15	Vision-language
QwenLM/Qwen3.8	4,115	Sep 15	3.8 series
QwenLM/Qwen-MM-Plugins	2,817	Sep 15	Multimodal agent plugins
QwenLM/FlashQLA	702	Sep 15	New — linear-attention kernels (TileLang)
QwenLM/Qwen-Drive-1.0	413	Sep 15	Autonomous-driving VLM
QwenLM/Qwen-RobotManip	167	Sep 15	Robot manipulation