In short
- Kimi-K3.1-Pro tops Hugging Face trending (1,589 likes, 22,571 downloads on 2 November) but has a high like-to-download ratio — the model to watch, not necessarily to run this week.
The short answer: what's worth running this week
The best LLM right now, if you're asking the question most people ask: Kimi-K3.1-Pro, Moonshot's newest release, which topped Hugging Face's trending page on 2 November with 1,589 likes and 22,571 downloads. But read the fine print before you clear disk space — the trending page lists it as image-text-to-text, so check what you actually need before grabbing it.
What people are actually downloading is something else entirely: the DeepSeek-V4.1-Flash GGUF quants. ISTA-DASLab's Coder GGUF has 243,190 downloads — about ten times Kimi's figure. AtomicChat's Uncensored GGUF has 117,016 — about five times.
Our verdict in one line: Kimi-K3.1-Pro if you want the newest frontier-ish open model and have the hardware for it; DeepSeek-V4.1-Flash quants if you want something that runs on a gaming GPU today.
This is our fortnightly "what's trending and what it costs" column: what's hot on Hugging Face, what it takes to run locally, and what it costs to call rather than host.
What's trending on Hugging Face right now
Five models made Hugging Face's trending page between 23 October and 2 November:
| Model | Trending | Likes | Downloads |
|---|---|---|---|
| Kimi-K3.1-Pro | 2 Nov | 1,589 | 22,571 |
| DeepSeek-V4.1-Flash Coder GGUF (ISTA-DASLab) | 2 Nov | 189 | 243,190 |
| DeepSeek-V4.1-Flash Uncensored GGUF (AtomicChat) | 28 Oct | 359 | 117,016 |
| DeepSeek-V4.1-Flash obliterated (OBLITERATUS) | 23 Oct | 270 | 3,835 |
| DeepSeek-V4-Pro-Qwen3.5-9B heretic (xtools-at) | 30 Oct | 281 | 1,369 |
The pattern is hard to miss: three of five are DeepSeek-V4.1-Flash variants, and three of five are community de-aligned or uncensored re-releases. Trending is not the same as "best" — it measures downloads and excitement, not quality.
Note the download gaps. The two GGUF quant packs sit at 243,190 and 117,016 downloads; the obliterated variant managed 3,835 and the heretic 9B just 1,369. The split isn't random: GGUF is the format llama.cpp, Ollama, LM Studio and koboldcpp all read — the tools most people already have installed — while the non-GGUF builds need more work to run. If you're weighing up which engine to install, our LM Studio vs Ollama comparison covers the two most popular options.
One honest caveat before we go further: we can't independently benchmark Kimi-K3.1-Pro or the Flash variants this week. The likes and downloads above are Hugging Face's own trending numbers, and any quality claims below are flagged as vendor- or community-reported.
Kimi-K3.1-Pro: the headline model, and the catches
Moonshot's Kimi-K3.1-Pro is the model getting the attention, and the numbers tell an interesting story. It's listed as image-text-to-text, with 1,589 likes against 22,571 downloads.
That like-to-download ratio is unusually high. Compare: the Coder GGUF has 189 likes and 243,190 downloads — people run it and move on. Kimi has over eight times the likes but a tenth of the downloads. That ratio suggests something big — a model lots of people are admiring and relatively few are running. Until a size, parameter count or quant is confirmed, assume it needs serious VRAM and plan accordingly.
What that means practically: if your hardware can't hold it, the honest answer to "best LLM right now" for you is something smaller — and that's fine. The Flash quants below exist precisely for that case.
One tagging note: the trending page lists both Kimi and the ISTA-DASLab Coder GGUF as image-text-to-text, so we wouldn't lean too hard on modality to separate them — it may just be how the repos are tagged. What we can say from the numbers is that Kimi is the attention magnet and the Flash quants are the workhorses.
The pattern to watch: community quants. Big open releases this year have generally picked up GGUF and MLX quant packs from the community once demand appears, and that's the most likely route to Kimi running on consumer hardware. If a Q4 quant lands and holds quality, it becomes runnable for far more people overnight. Until then, it's the model to watch, not necessarily the model to run this week.
DeepSeek-V4.1-Flash quants: what people actually download
The real story of this fortnight is the Flash quants.
ISTA-DASLab's DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF: 243,190 downloads, 189 likes — the most-downloaded of the five models in this window by a factor of two. It's a coder-focused quant pack. If you're doing local coding work, this is what the crowd has voted for.
AtomicChat's Uncensored GGUF: 117,016 downloads, 359 likes. The uncensored and de-aligned re-releases are a permanent fixture of trending — people want models without refusals for roleplay, writing and research, and they download rather than debate. We'll treat that fact the same way we treat any other download statistic: as data about what people run.
The trade-offs, which we always admit: de-aligned re-releases carry no publisher guarantees. Quality can drift from the original checkpoint, and you're trusting a community uploader you've probably never heard of — check the repo card for what was actually changed. Fine for a hobby. Think twice before wiring one into a product — you're taking on responsibility the original publisher refused to.
The smaller builds — OBLITERATUS's obliterated version (3,835 downloads) and xtools-at's heretic 9B (1,369) — exist but haven't caught on. Note these aren't the same kind of thing as the GGUF packs: "obliterated" and "heretic" are de-alignment methods applied to a checkpoint, not quantisation formats, so they're not competing on the same axis. What the numbers do show is that the two GGUF repos — the ones with "GSQ-RCO" in the name, whatever that string denotes — account for over 360,000 downloads between them, while the non-GGUF de-aligned builds sit in the low thousands.
Why GGUF matters, for anyone new here: it's the single format that llama.cpp, Ollama, LM Studio and koboldcpp all read natively. That's exactly why these quants out-download the other builds — they're the version that runs in the software you already have.
Which one should you run?
| Model | What it is | Likes / Downloads | Best for | Our verdict |
|---|---|---|---|---|
| Kimi-K3.1-Pro | Listed image-text-to-text | 1,589 / 22,571 | Newest frontier open model | Watch; run it if you have the VRAM |
| Flash Coder GGUF (ISTA-DASLab) | Coder quant pack | 189 / 243,190 | Local coding on a gaming GPU | The one we'd actually host |
| Flash Uncensored GGUF (AtomicChat) | De-aligned chat | 359 / 117,016 | Roleplay and unfiltered writing | Run locally, know the trade-offs |
| Flash obliterated (OBLITERATUS) | De-aligned build | 270 / 3,835 | Same idea, less traction | Skip; the GGUFs above are better supported |
| Heretic 9B (xtools-at) | 9B uncensored fine-tune | 281 / 1,369 | Small-GPU unfiltered chat | Niche; only if 9B is your ceiling |
How we'd host these, if we were pointing spare hardware at them: the Coder GGUF as a general-purpose endpoint — it has the demand, the format support and the broadest use case. Kimi only if big-hardware demand shows up, and only once a quant exists that ordinary machines can actually serve.
What it costs to call these instead of hosting them
Here's the cost angle nobody's listicle has: the hidden price of "best LLM right now" is the hardware. A GPU big enough for a frontier open model is a four-figure purchase, and for most people it sits idle most of the day. Per-token pricing flips that — you pay only for what you generate.
For context on the cheap end of open models: chat starts from £0.009 per 1M output tokens on Inpherio's marketplace (qwen2.5:0.5b and smollm2:360m), and 116 open models are callable through one OpenAI-compatible API right now — see the live prices on the marketplace.
To be clear: Kimi-K3.1-Pro and the Flash quants aren't on the marketplace as of today. That's the general cost picture, not a claim you can call them for pennies. For coding work in the meantime, qwen2.5-coder:7b is live from £0.063 per 1M output tokens.
If you own the hardware instead: hosts on Inpherio keep everything except a 15% platform fee, and new hosts get 12 months fee-free (up to £250 of earnings) — details at /providers.
The natural next step if you'd rather test a model for 20p than buy a GPU: the API quick start at /developers.
How we'll know if the verdicts change
Each verdict above has a signal that would flip it:
- Kimi-K3.1-Pro: watch for community quants (GGUF/MLX) and the first independent benchmarks. If a Q4 quant lands and holds quality, it becomes runnable for far more people — and our "watch, don't run" verdict changes.
- DeepSeek-V4.1-Flash: watch whether the GGUF quants hold up against independent evals, and whether the uncensored variants stay stable over longer conversations. De-aligned builds sometimes degrade in ways the original never did.
- The standing caveat we apply to every model in this column: trending measures attention, likes measure enthusiasm, downloads measure runnability — none of them measure quality. That's why we link the source numbers every time.
Next column runs in a fortnight. If a Kimi quant or an independent benchmark lands sooner, we'll cover it separately.
Questions people ask
What is the best LLM right now (November 2026)?
If you want the newest open model and have serious hardware, Kimi-K3.1-Pro tops Hugging Face trending. If you want something that runs on a gaming GPU today, the DeepSeek-V4.1-Flash GGUF quants are what most people are actually downloading.
Models in this post — live on Inpherio
Have a GPU, a Mac or a spare PC?
Let it earn while you sleep.
Run open models for people who need them. You set the price and keep everything but the 15% fee.
Become a hostBuilding with AI?
Every open model, one API.
OpenAI-compatible. Priced by the people who run the hardware.
Start free See the marketplaceRead “Best LLM right now: what's trending, what runs, what we'd pick (Nov 2026)” at https://inpherio.co.uk/blog/best-llm-right-now-what-s-trending-what-runs-what-we-d-pick-nov-2026