llama.cpp Adds MTMD Vision, Ollama 0.33 RC, and Qwen3.8 Community Forks
llama.cpp extends native multimodal coverage with dots3-note vision+audio, Ollama ships 0.33.0-rc2 with Claude Desktop integration, and Qwen3.8-27B drives heavy community quantization activity.
Top Stories
1. llama.cpp: MTMD Multimodal Subsystem Grows (PR #27524)
PR #27524 (“mtmd: support dots3-note vision+audio”) extends the MTMD multimodal subsystem with native dots3-note vision+audio support.
- What changed: MTMD now handles
dots3-notemodels natively, feeding images and audio directly into inference without external preprocessing. - Why it matters: Native multimodal in the reference engine means downstream tools (Ollama, LM Studio) inherit it without patching.
- Status: Merged PR, part of ongoing MTMD work.
- Source: PR #27524
2. Ollama 0.33.0-rc2: Claude Desktop Integration
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 and reworked prefill caching (fixes hang on agent-client prefill cancel).
- Status: RC; stable v0.33.0 not yet out.
- Source: Ollama v0.33.0-rc2
3. Qwen3.8-27B: Community Forks Are the Real News
Qwen/Qwen3.8-27B is #3 on HF trending (12,423 likes, 2.6M dl). The story is the ecosystem: uncensored/abliterated and quantized variants (GGUF, MLX, NVFP4) fill the top 100, led by unsloth GGUF at 7M+ downloads.
- VRAM (analytical): Q4_K_M ~16GB, fits RTX 3090 24GB; FP8 ~27GB needs 2x24GB or 48GB card.
- Source: HF Qwen3.8-27B | unsloth GGUF
Model Trends
| Model | Likes | Type | VRAM (est.) | Fit |
|---|---|---|---|---|
| FLUX.1-dev | 14,231 | Text-to-Image | ~12GB FP8 | RTX 3090 24GB |
| 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, excluding large KV cache. DeepSeek-R1 and Kimi-K3 are not single-consumer-GPU models.
Engine Updates
- llama.cpp PR #27524 (merged): MTMD dots3-note vision+audio.
- llama.cpp PR #27608 (merged): Metal/CUDA virtual device naming.
- llama.cpp PR #27614 (merged): Fix multi-GPU server tests.
- vLLM v0.27.1 (Aug 11): Quantized DSpark Markov heads patch.
- SGLang v0.5.18 (Aug 22): 710-PR release, adds Muse Glimmer + Intern-S2-Mobius.
Community
The Qwen3.8-27B fork ecosystem is the notable story: dozens of uncensored/quantized variants in the top 100, unsloth GGUF at 7M+ downloads. Self-hosters treat official weights as a base to re-quantize, not a finished product.
What Changed Since Last Scan
Delta vs 2026-08-21:
- llama.cpp: PR #27524 merged (MTMD dots3-note vision+audio)
- Ollama: v0.33.0-rc2 (Claude Desktop, Apps view, prefill caching)
- SGLang: v0.5.18 (Muse Glimmer, Intern-S2-Mobius)
- Qwen3.8-27B: 12,423 likes; community variants dominating forks
- No new official base models
Sources
Last updated: 2026-08-22