AI
AI Updates
Back to all posts
·
ai-updatesdaily-briefinghuggingfaceollamaqwen

Nvidia's $13B Hugging Face Talks, Ollama v0.33.1, and Flash-Next's Trending Climb

Nvidia is reportedly in talks to acquire Hugging Face for ~$13B (1,408 points on HN), Ollama ships v0.33.1 with MLX support for Qwen3.8-Flash-Next, and the new Flash-Next model climbs to 3,847 likes on HF trending.

A
AI Updates

Today’s story is structural, not a model drop: a 1,408-point Hacker News thread on reports that Nvidia is in talks to acquire Hugging Face for roughly $13B — the open-weights distribution layer now a strategic question. Meanwhile the engine loop we’ve tracked since 08-25 closes: the accelerated Ollama RC cadence ended in v0.33.1 stable, which ships MLX support for Qwen3.8-Flash-Next — the model now climbing HF trending at 3,847 likes.

Top Stories

1. Nvidia Reportedly in Talks to Acquire Hugging Face for ~$13B

Business Insider reports that Nvidia is in talks to buy Hugging Face for roughly $13B. To be clear on the status: this is reportedly in talks, not a signed deal — no filing and no announcement from either side in anything we tracked today. The HN thread (1,408 points, 626 comments) is the day’s story by an order of magnitude.

Why it matters for self-hosters: Hugging Face is the de-facto distribution layer for open weights. The trending page is where new model releases actually surface — today’s only new model, Qwen3.8-Flash-Next, is on it — and the Hub hosts the official weights for effectively every major open model. A $13B Nvidia acquisition would put the curation layer under the largest GPU vendor, and the open questions are the ones self-hosters care about:

  • Curation neutrality — does the trending page, which decides which open weights the community actually discovers, start to favor models that run best on NV hardware?
  • Competitor hosting — does Nvidia host and promote models from AMD, Intel, and Apple with the same prominence as its own?
  • API coupling — the Hub’s inference API is the default first stop before anyone reaches for a local runtime; an ownership change raises where that traffic lands.

None of that is decided or even confirmed. But the thread’s size is itself the signal: the community is already treating the open-weights distribution layer as a strategic asset, not a neutral utility.

2. Ollama v0.33.1: The RC Cadence Closes

Ollama v0.33.1 published 18:09 UTC on Aug 26 — five days after v0.33.0 stable cut on Aug 21, the same evening as the rc2/rc3 double we tracked on 08-25. The cadence call — two candidates in one day means stable is close — held: stable shipped that night, and this patch followed within the week.

What actually shipped in v0.33.1:

  • MLX: Qwen3.8 Flash Next support — the headline. In our tracking, v0.33.1 is the first stable release to ship any runtime path for Flash-Next.
  • mlxrunner: structured output support — agent loops on local MLX models can now constrain output to a schema.
  • mlxrunner: avoid Metal GPU timeouts when loading models from slow storage — a real fix for anyone loading large models from a NAS or a slow SSD.
  • MLX and llama.cpp dependency updates, a cmake idempotency patch, and first contributions from new contributors (PR #17948).

Why it matters: Flash-Next support lands in Ollama’s MLX runtime before the llama.cpp GGUF path — those PRs are still open (below). For Apple Silicon self-hosters, v0.33.1 is the first stable way to run the day’s trending model.

Qwen3.8-Flash-Next is the only new model in today’s trending delta, already at 3,847 likes (up from 3,044 in yesterday’s post — roughly 800 likes in a day). It’s an image-text-to-text model tagged qwen4_exp — an experimental artifact on the Qwen 4 line, which is what makes it worth watching: the first public surface of the next architecture generation, not just a speed variant of 3.8. The official FP8 variant (110 likes) is out for the datacenter path, and the model repo describes it as “the foundation model developed by Qwen Team, Alibaba Group.”

Engine support, where things actually stand:

The tracker roster rounds out the stack. Five new repos landed on the Qwen org tracker today: Qwen (21.7k★, the original), Qwen3 (27.6k★), Qwen-Agent (17.0k★), qwen-code-examples, and Qwen-MM-Plugins (2.8k★, “Make any agent harness multimodal-native”). The tracked roster now spans 10 repos — from the original Qwen repo through the Flash-Next foundation-model repo and qwen-code (27.4k★), the terminal coding agent. The full open-weights stack, in one tracker.

4. llama.cpp b10647–b10649: Three Nightlies in an Hour

llama.cpp cut three builds on 2026-08-27 (11:25 → 12:28 UTC):

  • b10647: args: add --video-* CLI arguments (#24318) — video-input arguments at the CLI level. b10647
  • b10648: model : simplify MiniMax-01 graph (#27790) — cleanup of the MiniMax-01 implementation. b10648
  • b10649: spec: Add benchmark-only synthetic speculative acceptance options (#27711) — synthetic speculative-decoding acceptance for benchmarks in llama-server/llama-cli. b10649

Compared to what came before: none of this is user-facing the way the 08-25 cycle was (no stable tag, no OOM fix). The cadence — three builds in an hour — plus the open Flash-Next PRs above is the signal: the tree is moving toward the next model generation while the 27B ecosystem settles.

Top 10 trending (likes / downloads):

Model Likes Downloads
FLUX.1-dev 14,265 670,366
DeepSeek-R1 13,589 3,892,061
Qwen3.8-27B 12,973 3,457,687
Kimi-K3 11,032 2,829,554
SDXL-base-1.0 8,075 1,584,385
SD-v1-4 7,055 476,797
Kokoro-82M 6,754 12,039,437
Llama-3.1-8B-Instruct 6,675 6,166,772
Meta-Llama-3-8B 6,633 1,066,159
whisper-large-v3 6,190 4,520,368

Three data points worth a look:

  • Qwen3.8-27B-FP8 now has more downloads (3.97M) than the base model (3.46M) — the pattern we flagged on 08-25 continues: datacenter users pull the FP8 weights and quantize on the fly.
  • Just off the top 10, Qwen3.8-Flash-Next (3,847 likes) is the only new face on the list — the day’s model story is a climb, not a launch-day spike.
  • all-MiniLM-L6-v2 sits at 251M downloads — the invisible workhorse of embedding pipelines and the most-downloaded model on the entire trending page.

AI News

  • “CEO fired developers to make room for AI. Developers create open source AI CEO” — 746 points. OpenExecutive is the open-source response: an AI executive layer built by the developers who were displaced. HN thread. Why it matters here: a concrete case of the “AI replaces the engineering org” dynamic, and the community’s counter-move — open-sourcing the executive function rather than accepting the vendor one.

What Changed Since Last Scan

Delta vs 2026-08-26 — 13 new items:

  • New releases: llama.cpp b10647, b10648, b10649 · Ollama v0.33.1
  • New HF models: Qwen/Qwen3.8-Flash-Next (3,847 likes)
  • New Qwen org repos on the tracker: Qwen, Qwen3, Qwen-Agent, qwen-code-examples, Qwen-MM-Plugins
  • New HN stories: 3

Summary: the open-weights distribution layer became the day’s story (reportedly Nvidia/HF talks, 1,408 points), the Ollama RC cadence we tracked for two days closed with v0.33.1 shipping MLX support for Qwen3.8-Flash-Next, and Flash-Next itself is the fastest riser on trending — with the llama.cpp GGUF path still open.

Sources

Last updated: 2026-08-27