AI
AI Updates
Back to all posts
·
ai-updatesdaily-briefingollamaqwenllama-cpp

Ollama Models Move Into ChatGPT Desktop — and Qwen's Flash-Next Signals Qwen4

Ollama v0.34.0-rc1 lets you run Ollama models directly inside ChatGPT Desktop on macOS, bridging local open models into OpenAI's consumer client — while Qwen's qwen4_exp-tagged Flash-Next climbs the charts and llama.cpp ships a three-build day (M2 Max FA-vec, CMake UI embeds, --log-jsonl).

A
AI Updates

🔥 Top Stories

1. Ollama models now run inside ChatGPT Desktop (v0.34.0-rc1)

The story of the day: Ollama v0.34.0-rc1 (published 2026-09-05, release candidate) lets you use Ollama models directly in ChatGPT Desktop — “so you can keep your existing workflow while running open models,” with setup from the Ollama app on macOS.

Why it matters: this is the first concrete bridge between the two worlds of 2026 local AI — the dominant open-model runner (Ollama) and OpenAI’s flagship consumer client (ChatGPT Desktop). Until now, “local open model” and “ChatGPT” were separate products with separate UIs. Now the same ChatGPT Desktop interface can be pointed at your local weights. For self-hosters the pitch is direct: keep the workflow you already have, but run the model on your own GPU instead of paying per token. It also reinforces the “OpenAI-compatible endpoint” as the interop layer — the same RC adds OpenAI-compatible client tool search and response compaction, improved structured-output performance on Apple Silicon, and fixes images through compacted responses.

Caveats: it’s an RC (v0.34.0-rc1), not a stable, so expect v0.34.0 shortly, and the setup is macOS-only for now.

Ollama v0.34.0-rc1

2. llama.cpp’s three-build day: b10821 → b10823

Continuing the high-velocity cadence from yesterday (b10817–19), llama.cpp shipped three more builds on 2026-09-06, each a targeted change:

  • b10821 — Metal: add the remaining FA-vec tunings for M2 Max (#28458). The direct continuation of the flash-attention-vectorized work that landed for M3 Max in 0.4.0 (09-04) and the Metal memory-leak fix in b10819 (09-05) — now closing the loop on M2 Max for the Q4_0/Q4_1/Q5_0/Q5_1 rows. If you run Qwen3.8 on an M2 Max Mac, this is the perf one.
  • b10822 — build/UI: embed UI assets directly with CMake (#28445) — drops the build-time C++ helper and the external gzip dependency, simplifying cross-compilation and packaging.
  • b10823 — common: add --log-jsonl (#28437) — structured JSON-lines logging, a small but real observability win for anyone piping llama.cpp into a log shipper or dashboard.

In flight this scan: merged — a CUDA race fix in mmid/mmf caught by --tool=racecheck on MUL_MAT_ID (#28475) and a grammar max-repetition threshold fix (#28469). Open and worth watching — a CPU fallback when a KV/recurrent-state tensor exceeds the device limit (Vulkan’s maxStorageBufferRange, #28465), a genuine robustness win for AMD/Vulkan users, and a tiled TQ2_0 × Q8_K CPU prefill kernel in llamafile_sgemm (#28452).

b10823 · b10822 · b10821

3. Qwen’s next-gen signal: Flash-Next climbs, the quant ring keeps growing

The Qwen tracker’s signal today is a tag: qwen4_exp. Qwen3.8-Flash-Next — an official Qwen model, image-text-to-text (multimodal) and conversational, tagged qwen4_exp — is now at 4,932 likes / 432,966 downloads (up from 4,898 / ~401K yesterday), and it just gained its own official GitHub repo (QwenLM/Qwen3.8-Flash-Next, 330★). The tag reads as an early next-generation (Qwen4) experimental build; “Flash-Next” slots it into Qwen’s fast/efficient tier. An FP8 sibling (Qwen3.8-Flash-Next-FP8) is already live for low-VRAM serving.

Around the Qwen3.8-27B workhorse, the community quant ring (yesterday’s lead story) keeps compounding:

  • unsloth/Qwen3.8-27B-GGUF is now at 10.31M downloads (3,563 likes) vs. the official 6.19M — the gap has widened to ~1.67× (unsloth GGUF).
  • A fresh fine-tune/merge variant entered the tracker today: DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion…-GGUF (235 likes / 211K dl) — the “Cold-Fusion / Heretic / MTP” personality-tuning subgenre of the Qwen3.8 ecosystem.
  • The dense 2.4T-A95B MoE flagship (1,203 likes) and Qwen3-VL (19,905★) round out the official family.

Why it matters: the qwen4_exp tag is the first concrete data point that Qwen4 is landing in the wild, and Flash-Next — the “fast” multimodal tier — is the model most likely to be the default local chat/VLM of the next generation. Meanwhile the 10.3M-vs-6.19M gap confirms the distribution channel has fully moved to community repos.

Qwen3.8-Flash-Next · Qwen3.8-27B

📊 Model Trends

HF trending is stable at the top, but two things stand out this scan: OpenAI’s open-weight gpt-oss family quietly holds two top-20 slots with ~11.7M combined downloads, and all-MiniLM-L6-v2 — a plain embedding model — is the true download king (253M).

Rank Model Likes Downloads Role VRAM (Q4_K_M)
1 black-forest-labs/FLUX.1-dev 14,481 760,632 Text-to-image fits 24GB ✓
2 Qwen/Qwen3.8-27B 14,074 6.19M Chat / VLM fits 24GB ✓
3 deepseek-ai/DeepSeek-R1 13,613 789,555 Reasoning (MoE) full needs 8+ 3090s ✗ (32B distill ✓)
4 moonshotai/Kimi-K3 11,214 2.46M Chat / VLM ~16–24GB ✓
5 stabilityai/stable-diffusion-xl-base-1.0 8,113 1.77M Text-to-image fits 12GB ✓
6 CompVis/stable-diffusion-v1-4 7,057 537,236 Text-to-image fits 8GB ✓
7 hexgrad/Kokoro-82M 6,822 11.5M TTS trivial (82M) ✓
8 meta-llama/Llama-3.1-8B-Instruct 6,802 5.63M Chat fits 8–12GB ✓
9 meta-llama/Meta-Llama-3-8B 6,644 356,149 Base fits 8–12GB ✓
10 openai/whisper-large-v3 6,243 5.09M ASR fits 8GB ✓
12 sentence-transformers/all-MiniLM-L6-v2 5,561 253M Embeddings (RAG default) trivial (~90M) ✓
15 openai/gpt-oss-120b 5,158 5.23M MoE chat (vLLM-native) needs 3×24GB / 2×48GB ✗

Analysis:

  • OpenAI’s open weights are the quiet heavyweight. The gpt-oss 120b + 20b pair has ~11.75M combined downloads (5.23M + 6.52M), and both are tagged vllm — vLLM-native MoE models you can actually self-host. Until now, “an OpenAI model you run on your own GPU” basically didn’t exist in this list. That’s the notable shift. gpt-oss-120b · gpt-oss-20b
  • all-MiniLM-L6-v2 is the real download king (253M) — it’s the default sentence-embedding model under most RAG stacks, so it quietly dwarfs every generative model. Infrastructure, not a “trend.” all-MiniLM-L6-v2
  • Kokoro-82M (11.5M) is still the top generative model by downloads — the “tiny model everyone ships” TTS pattern holds.
  • DeepSeek-R1 (13,613) and Qwen3.8-27B (14,074) hold the #2/#3 local-reasoning photo finish; Kimi-K3 (2.46M dl) remains the strongest local VLM and the exact target of vLLM 0.28.0’s perf push.
  • Image gen still holds 3 of the top 6 (FLUX.1-dev, SDXL, SD v1-4); Z-Image-Turbo (Tongyi-MAI, 677K dl) is the new challenger in that slot.

HF explore

⚙️ Engine Updates

Ollama v0.34.0-rc1 (2026-09-05) — the headline: run Ollama models inside ChatGPT Desktop (macOS) + OpenAI-compatible tool search/response compaction + Apple Silicon structured-output. RC, stable v0.34.0 to follow. Release

llama.cpp — b10821 / b10822 / b10823 (2026-09-06) — one line per build (detail in Top Story #2):

  • Apple (Metal): b10821 closes out the M2 Max FA-vec tunings. Update if you run Qwen3.8 on M2 Max.
  • Build/UI: b10822 embeds UI assets via CMake (simpler cross-compile).
  • Observability: b10823 adds --log-jsonl.

b10823

SGLang v0.5.19 (2026-09-05) — carried from yesterday: native Qwen3.8 (27B + 2.4T-A95B MoE), 786 PRs / 214 contributors. Still the reason to upgrade if you serve Qwen3.8 at scale. Release

vLLM v0.28.0 (2026-08-26) — carried: the Kimi-K3 performance push (Decode Context Parallel, fused FlashKDA decode/prefill, SiTU for MegaMoE, combined all-gathers 1.5–3×). The version to be on for Kimi-K3 (HF #4). Release

TGI v3.3.7 / mlc-llm — both stale (TGI in explicit “maintenance mode”; mlc-llm’s latest release is from 2023). No action.

📰 AI News (Hacker News)

The AI feed is thin on launches and heavy on agent meta — the dominant thread is still the OpenAI agent story, now its second day leading:

  • Discovery of a new OpenAI agent message board (2,213 pts / 1,560 comments) — up from 1,802 yesterday, still the clear feed leader. The “message board” where OpenAI’s agents post to each other; the discussion has moved to what agent-to-agent surfaces actually look like and who’s accountable when agents talk. Source
  • AI, Tools and Transformation (86 pts / 37) — Ben Evans (a16z) on how AI tools are reshaping work; an industry-analysis piece rather than a launch, useful as the “what does this all mean” counterweight to the model-news treadmill. Source
  • Don’t Use a gmail.com Address (27 pts / 6) — a low-signal privacy/security take, included only for completeness of the scan. Source

🔄 What Changed Since Last Scan (vs 2026-09-05)

  • Ollama: v0.34.0-rc1 — Ollama models now usable inside ChatGPT Desktop (macOS), plus OpenAI-compatible tool search/response compaction. Action: if you’re on macOS and want local models in the ChatGPT workflow, try the RC.
  • llama.cpp: three new builds — b10821 (M2 Max FA-vec), b10822 (CMake UI embed), b10823 (--log-jsonl); in-flight: CUDA mmid/mmf race fix + Vulkan CPU-fallback. Action: update on Apple; watch #28465 if you run Vulkan/AMD.
  • Qwen: Qwen3.8-Flash-Next climbs to 4,932 likes and gains an official repo; the qwen4_exp tag is the next-gen signal. Qwen3.8-27B community GGUF widens to 10.31M vs 6.19M official. Action: watch Flash-Next as the likely Qwen4 fast model; keep pulling community quants, not raw weights.
  • HF trends: gpt-oss family holds two top-20 slots (~11.7M combined dl); all-MiniLM-L6-v2 is the true download king (253M); Z-Image-Turbo is the new image-gen challenger.
  • HN: two new stories (Ben Evans 86, gmail 27); the OpenAI agent message board (2,213) continues to lead, up from 1,802.
  • No new official HF dense flagships day-over-day — the action is in the Ollama bridge, the Qwen qwen4_exp signal, and the llama.cpp backend cadence.

Sources

Generated on 2026-09-06