Ollama Enters Claude Desktop, llama.cpp Ships MTMD Video Fix, Kimi K3 Serving Matures
Ollama 0.33.0-rc2 integrates local models into Claude Desktop, llama.cpp ships MTMD video and GLM conversion fixes, and Kimi K3 (2.8T LatentMoE) gains full serving support in vLLM and SGLang.
Top Stories
1. Ollama v0.33.0-rc2: Local Models Inside Claude Desktop
v0.33.0-rc2 (Aug 21) wires local models into Claude Desktop: toggle Ollama models on/off inside Claude, pick which local model handles a task. New Apps view. Reworked prefill caching fixes a hang where agent clients cancel long prefills.
- Why it matters: Concrete step toward hybrid local+frontier workflows.
- Status: RC; stable v0.33.0 not out yet.
- Source: Ollama v0.33.0-rc2
2. llama.cpp: MTMD Video Fix, GLM Fix, Open Hexagon NPU PR
Three builds on Aug 24:
- b10608 (12:55): “mtmd: video: fix moov atom at end of file (#27596)” - fixes video files with trailing moov atom.
- b10606 (12:31): “ggml: fix ggml_clamp (#27644)” - corrects clamp operation.
- b10605 (08:57): “mamba2: Flatten in/out projections to dispatch GEMM instead of GEMV (#27513)” - Mamba2 perf.
- Merged PR #27655: “convert: fix GLM regression in index_tensors” (fixes regression from #26534).
- Open PR #27642 (NOT merged): “hexagon: add alternative Hexagon NPU backend” - FastRPC-based ggml-hexagon for Qualcomm NPU (Android/Windows-on-Snapdragon). Under review; do not plan around it.
- Source: releases | PR #27655 | PR #27642
3. Kimi K3 Gets Full Serving Support (vLLM + SGLang)
moonshotai/Kimi-K3 - 2.8T multimodal LatentMoE (896 experts, top-16, 3584-dim latent, 1M ctx, MXFP4) - is #4 on HF trending (10,956 likes). Both engines now serve it:
- vLLM v0.27.0 (Aug 10): full Kimi K3 stack (kernels, Python+Rust frontends, DeepGEMM, DSpark AR fusion, shardable shared expert).
- SGLang v0.5.17 (Aug 8): day-0 support (DCP, DSpark, chunked-prefill PP, KDA-aware prefix caching, HiCache L2).
- Why it matters: 2.8T servable out-of-the-box in both engines = datacenter-scale, but signals frontier MoE serving becoming commodity.
- Source: vLLM | SGLang | Kimi-K3
4. Qwen3.8-27B: Fork Ecosystem Deepens
Qwen/Qwen3.8-27B #3 on HF (12,423 likes, 2.6M dl). unsloth GGUF crossed 7M downloads. New orcarouter/Qwen3.8-27B-Uncensored in this cycle’s delta.
- VRAM (analytical): Q4_K_M ~16GB, fits RTX 3090 24GB; FP8 ~27GB needs 2x24GB or 48GB.
- Source: Qwen3.8-27B | unsloth
Model Trends
| Model | Likes | Type | VRAM (est.) | Fit |
|---|---|---|---|---|
| FLUX.1-dev | 14,231 | Text-to-Image | ~12GB FP8 | RTX 3090/4090 |
| DeepSeek-R1 | 13,584 | Reasoning 671B MoE | Multi-GPU | Datacenter |
| Qwen3.8-27B | 12,423 | General 27B dense | ~16GB Q4_K_M | RTX 3090 24GB |
| Kimi-K3 | 10,956 | Multimodal 2.8T LatentMoE | MXFP4 multi-node | Datacenter |
| SDXL base-1.0 | 8,068 | Text-to-Image | ~8-16GB | RTX 3090 24GB |
| Llama-3.1-8B-Instruct | 6,658 | General 8B | ~5GB Q4 | Mac/8GB |
VRAM = analytical estimates. DeepSeek-R1 (671B MoE) and Kimi-K3 (2.8T) are not single-consumer-GPU models. DeepSeek-V4-Pro also trending (~5,475 likes) - trending, not a new release.
Engine Updates
- Ollama v0.33.0-rc2 (Aug 21): Claude Desktop, Apps view, prefill caching.
- vLLM v0.27.0 (Aug 10): Kimi K3 full stack. v0.27.1 (Aug 11): DSpark Markov heads patch.
- SGLang v0.5.17 (Aug 8): Kimi K3 day-0. v0.5.18 (Aug 22): Muse Glimmer + Intern-S2-Mobius.
- llama.cpp: b10605/b10606/b10608. Merged: #27655 (GLM), #27614 (multi-GPU tests), #27608 (Metal/CUDA naming). Open: #27642 (Hexagon, NOT merged).
AI News
- FT: Anthropic’s best AI model struggles as cheaper tools thrive (HN 604 pts). Source: HN 49411102
- “My agent.md to improve LLM-assisted code quality” (HN 336 pts). Source: HN 49410932
- Paul Graham: “I were 17, I’d learn how to build LLMs from scratch” (HN 251 pts). Source: HN 49412396
- “I built a low-latency AI companion that plays Skyrim with me” (HN 221 pts). Source: HN 49413561
What Changed Since Last Scan
Delta vs 2026-08-23:
- New llama.cpp releases: b10605, b10606, b10608
- New Qwen model: orcarouter/Qwen3.8-27B-Uncensored (142 likes)
- New Qwen repos: Qwen-RobotNav (154), Qwen-Image (8,257), Qwen3-ASR (3,410)
- New HN stories: 11
- New HF models: none
Summary: No frontier base-model launch. Defined by serving and integration: Kimi K3 servable across engines, Ollama in Claude Desktop, Qwen fork ecosystem deepening.
Sources
Last updated: 2026-08-24