Model Intelligence — 2026-06-17 (Updated)

🔥 Top Stories

1. Ollama v0.30.10 Goes Stable — Cohere2MoE Support Arrives

Ollama just shipped v0.30.10 stable (June 17, 4:22 PM UTC), upgrading from the RC released earlier today. The headline feature: Cohere2MoE model architecture support via PR #16670, plus a llama.cpp bump to b9672.

Why it matters: MoE (Mixture-of-Experts) models are the architecture of choice for efficient large-scale inference. Cohere's MoE designs split computation across specialist sub-networks, enabling 50-100B parameter models to run with the active compute budget of a 10-20B dense model. Ollama's native support means these models become one ollama run command away.

2. llama.cpp b9692: Server API + Metal Improvements Continue

Four new builds landed since the morning scan — b9689, b9690, b9691, b9692 — continuing a blistering development pace. Standout changes:

• b9692 — llava_uhd batch dim fix: llava_uhd should no longer use batch dim, fixing multi-image processing in the UHD variant of LLaVA vision models.
• b9691 — Power11 backend (conditional on compiler support): IBM Power11 processors are getting native GGML acceleration. If you're running on enterprise Power hardware, this unlocks hardware-accelerated inference without GPU offload.
• b9690 — Metal rope_back operator: Apple Silicon users gain optimized backward RoPE rotation support. This is part of the speculative decoding stack — the rope_back operator enables efficient multi-token prediction evaluation paths.
• b9689 — Metal f16/bf16 concat operator: Extended Metal backend support for half-precision concatenation, improving performance for models with multi-head attention that use concat operations.

Server-side update: b9688 introduced a model management API for the llama.cpp router server, adding programmatic model control capabilities.

Combined with yesterday's SYCL improvements (VRAM overcommit, fp16 ops, MoE prefill fix), llama.cpp is now shipping meaningful backend improvements across every major GPU vendor simultaneously: NVIDIA (CUDA), Intel (SYCL), AMD (Vulkan/OpenCL), Apple (Metal), and IBM (Power11).

3. DeepSeek-V4-Pro Accelerates — Now at 4,926 Likes

DeepSeek-V4-Pro moved from 4,915 → 4,926 likes (+11 in the last scan), maintaining its #13 trending position. The acceleration suggests growing community interest in DeepSeek's next-gen MoE architecture.

The V4-Pro architecture is a hybrid MoE design that excels at both reasoning and general tasks. With vLLM v0.23.0 providing production-grade V4 support and llama.cpp's NVFP4 hardening from yesterday, the model is now accessible across both throughput-focused (vLLM) and latency-focused (llama.cpp) serving stacks.

📊 Model Trends

HuggingFace Trending (Top 15)

Signal: The top-10 is remarkably stable. The only movement is in the 11-15 band where DeepSeek-V4-Pro continues its steady march past gpt-oss-120b. Kokoro-82M (+3) and all-MiniLM-L6-v2 (+4) show utility models (TTS, embeddings) maintaining consistent growth.

Qwen Ecosystem

|| Model | Likes | 24h Δ | VRAM Fit | |-------|-------|-------|----------| | Qwen/QwQ-32B | 2,931 | — | RTX 3090 @ Q4 (~19GB) | | Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled | 2,882 | — | RTX 3090 @ Q4 (~17GB) | | Qwen/Qwen-Image | 2,511 | — | Multi-modal | | Qwen/Qwen-Image-Edit | 2,425 | — | Multi-modal | | Qwen/Qwen3.6-35B-A3B | 2,157 | +6 | RTX 3090 @ Q3 MoE (~12GB) | | Qwen/Qwen2.5-Coder-32B-Instruct | 2,045 | — | RTX 3090 @ Q4 (~19GB) | | Qwen3.6-35B-A3B-Uncensored | 1,941 | +13 | RTX 3090 @ Q3 MoE (~12GB) | | Qwen/Qwen2.5-Omni-7B | 1,909 | — | RTX 3060 @ Q4 (~4GB) ✅ | | Qwen/Qwen3.6-27B | 1,741 | new entry | RTX 3090 @ Q4 (~16GB) |

Note: The uncensored variant continues outpacing the official (+13 vs +3), a pattern seen over multiple days. Community demand for unfiltered Qwen3.6 reasoning is real.

Gemma Ecosystem

|| Model | Likes | 24h Δ | VRAM (Q4_K_M) | |-------|-------|-------|--------------| | google/gemma-7b | 3,359 | +1 | ~4.5GB ✅ | | google/gemma-4-31B-it | 3,014 | +1 | ~17GB | | google/gemma-3-27b-it | 1,980 | — | ~15GB | | google/gemma-3n-E4B-it-litert-preview | 1,485 | — | ~2.4GB ✅ | | google/gemma-2-2b-it | 1,392 | — | ~1.3GB ✅ | | google/gemma-3-4b-it | 1,371 | — | ~2.3GB ✅ | | google/gemma-4-E4B-it | 1,257 | +1 | ~2.4GB ✅ | | google/gemma-7b-it | 1,247 | — | ~4.5GB ✅ | | google/gemma-2b | 1,194 | — | ~1.1GB ✅ | | google/gemma-4-26B-A4B-it | 1,155 | +4 | ~14GB |

New entry: A community GGUF build of Gemma 4 12B appears in the recent uploads — yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF. This is a specialized coding model variant already formatted for local inference.

⚙️ Engine Updates

llama.cpp — 14 Builds in 24h (b9673 → b9692)

Today's builds span b9673 through b9692 — a 24-hour sprint across multiple backend improvements:

Bottom line: If you have an Apple Silicon Mac, upgrade to b9692 for the Metal rope_back and concat improvements. If you have an Intel ARC GPU, the b9673-b9675 series is transformative. NVIDIA CUDA users get less direct benefit today but the Power11 addition signals broader hardware support.

Source: llama.cpp releases

Ollama — v0.30.10 (June 17, STABLE RELEASE)

Stable release: v0.30.10 graduated from RC to stable today. Cohere2MoE support is the primary payload, making MoE models accessible with ollama run.

Source: Ollama releases

vLLM — v0.23.0 (June 15, no new release)

Still the latest. This remains the most significant serving-stack release of the month: 408 commits, DeepSeek-V4 hardening, MRv2, Rust frontend, Gemma 4 Unified, multi-tier KV cache.

Source: vLLM releases

SGLang — v0.5.13 (June 13, no new release)

Still at v0.5.13. Nemotron 3 Ultra support added in this release. No new builds in 4 days.

Source: SGLang releases

📰 AI News (Hacker News)

The HN AI story set today is lighter than usual — only one AI-specific story passed the filter:

• [6 pts] "A Robot Is Sprinting Towards You: Do You Want It Running on Claude or Grok?" — An OpenRouter blog post exploring agent reliability through a robot-safety scenario. The low score suggests this is either very new or not resonating yet. Worth revisiting tomorrow.

From the broader HN conversation (carrying forward from yesterday's notable stories):

• [1,466 pts] "Running local models is good now" (Vicki Boykis) — Still dominating HN. The argument: quantization quality, hardware availability, and ecosystem maturity have converged to make local inference genuinely practical. Link
• [503 pts] "GLM-5.2 is the new leading open weights model on Artificial Analysis" — A competitive signal worth monitoring for next week's trends. Link

🔄 What Changed Since Earlier Today

The key update: Ollama v0.30.10 stable is the most significant new release since the morning scan. The llama.cpp b9688-b9692 batch adds meaningful Metal backend improvements for Apple Silicon users plus a server model management API. Model like counts show steady organic growth — no dramatic shifts. The Qwen3.6-27B entry into the top 10 signals growing interest in the 27B parameter tier.

🎯 Quick Recommendations for Your GPU

RTX 3060 (12GB):

• Qwen/Qwen2.5-Omni-7B at Q4_K_M (~4GB) — comfortable multi-modal option
• google/gemma-4-E4B-it at Q4_K_M (~2.4GB) — blazing fast chat
• google/gemma-3-4b-it at Q4_K_M (~2.3GB) — lightweight alternative

RTX 3090/4090 (24GB):

• Qwen3.6-35B-A3B at Q3_K_M (~11-13GB) — MoE efficiency, best reasoning in class
• google/gemma-4-26B-A4B-it at Q4_K_M (~14GB) — MoE, great instruction-following
• Qwen/QwQ-32B at Q4_K_M (~17GB) — strong reasoning
• deepseek-ai/DeepSeek-V4-Pro at Q4_K_M — excellent performance with vLLM

Intel ARC (B580 12GB / A770 16GB):

• Upgrade llama.cpp to b9692 immediately for SYCL fp16 + VRAM overcommit
• Qwen3.6-35B-A3B with GGML_SYCL_USM_SYSTEM=1 for MoE models that'd otherwise OOM

Apple Silicon (M-series):

• Upgrade llama.cpp to b9692 for Metal rope_back + concat f16/bf16
• All 8B-32B models benefit from the improved Metal backend operators

Model Intelligence brief generated 2026-06-17 by Hermes Agent.

Sources: HuggingFace API, llama.cpp releases, Ollama releases, vLLM releases, SGLang releases, Hacker News