LLMs

Jev Wants to Reshape the LLM: Is There Life Beyond Transformer?

Simon Willison flagged a new model called Jev, billed as a 'new shape' for LLMs. It hit Hacker News on September 21 with just 20 points and 3 comments — a quiet debut that could signal either a slow start or an under-the-radar opening move.

Font size

Jev: What It Is and Why It Made the HN Front Page

On September 21, Simon Willison's blog posted a short piece titled "Jev: A new shape for LLMs" — the phrase "new shape" sitting there unquoted, no question mark. Source: simonwillison.net/2026/Sep/21/jev/.

Jev's developers position it as a "new kind of LLM," pushing a path that diverges from today's mainstream architectures. Simon isn't the type to blog about mediocre releases — among independent commentators, he's one of the fastest at picking up new models. Mistral's early days, the various Llama variants, GPT-5 previews: he tends to be the first to weigh in. So if he's writing about it, Jev isn't another LoRA-tuned wrapper.

Same day, Jev hit the Hacker News front page. Score 20, comments 3 — a slightly odd ratio. On a normal HN front-page post, comment counts usually run 30–60% of the score, so a 20-point post would typically draw 8–15 comments. Three suggests either that lots of people clicked in but no one dared to speak up, or that actual viewership was thinner than the score implied. Both scenarios get unpacked below.

HN thread: item?id=49796843.

Where Is the "New Shape" Actually New?

Honestly, I can't write a definitive answer yet, because what's publicly available right now includes:

  • No benchmark numbers
  • No technical paper or blog post
  • No description of the training method
  • No open weights (as of Simon's post going live)

Simon's blog summary itself is one line: "A new LLM with a new shape." That kind of minimalist wording is unusual in the AI world — people usually try to cram "revolutionary" into the headline at all costs.

Based on the available clues, here are the directions I can guess at:

  1. Non-Transformer backbone — state space models (the Mamba line), linear attention variants, or some hybrid architecture.
  2. Non-autoregressive decoding — parallel generation, diffusion-style LLMs (like Mercury's approach to sandwiched generation).
  3. New topology — non-chain-of-thought, non-token-by-token, switching to graph or tree-structured generation.
  4. Complete rethink — some new memory mechanism replacing the KV cache.

I personally lean toward #1 or #4, since those are the directions where the most non-Transformer work has accumulated over the past two years. But this is speculation — Jev itself will have to confirm.

Why Nobody's Commenting — It's Not Lack of Interest, It's Lack of Material

The strange 20-points-to-3-comments ratio gets explained like this:

People looked. The community just has nothing to work with.

Comment activity on technical HN posts tracks strongly with two things:

  • Verifiable benchmarks: numbers on MMLU, HELM, HumanEval, LiveCodeBench — the moment those land, the comments ignite.
  • Readable code or papers: an arxiv link, a GitHub repo, or a demo URL.

Jev has none of the three. For an LLM claiming a "new shape," with not even a loss curve published, the experienced users in the comments stay silent — because anything they'd say is pure guesswork, and no one wants to look like a fool.

That also explains why there are still 3 comments at 20 points: those three are almost certainly low-effort curiosity posts ("Interesting," "Watching"), not real technical discussion.

When I followed Mamba's early days and the Mistral 7B leak period, the HN discussion curve looked identical: silence first, then an explosion the moment a repo or paper dropped.

Transformer Hegemony: What Became of the Challengers

Roll the clock back from late 2023 to early 2026, and at least these Transformer alternatives surfaced:

  • Mamba / Mamba-2 — state space models, linear complexity. Papers and open-source releases exist, but ecosystem support (Hugging Face, vLLM) lags far behind the Llama family.
  • RWKV — RNN + Transformer hybrid, emphasizing efficient long-context inference. A niche community runs it; major vendors haven't adopted it.
  • RetNet / TransNormer — multi-scale decay attention, marketed as a Transformer replacement. Microsoft-affiliated projects with strong academic visibility but almost no industrial deployment.
  • xLSTM — extending LSTM's memory mechanism. Proposed by Sepp Hochreiter's team in 2024; paper-level buzz.
  • Mercury / diffusion LLMs — diffusion-style generation, parallel token output. Inception Labs is working on it; commercial demos are live.

The common thread is painful: inference is genuinely cheaper, but the ecosystem is no match for Transformer.

What counts as ecosystem?

  • One-line loading via Hugging Face Transformers
  • Native support from inference frameworks like vLLM, TensorRT-LLM, llama.cpp
  • Tooling for community fine-tuning (LoRA, QLoRA)
  • Tutorials, blogs, Stack Overflow answers

That stack doesn't accumulate in a year or two. From Llama 2's release to vLLM stable support took more than a year. Mamba's vLLM support is still labeled experimental today.

So if Jev really is non-Transformer, the first problem it has to solve isn't "my model is better" — it's "my model can actually run."

How I'd Verify Whether Jev Is Real or Just a Pitch Deck

Here's my own practical checklist for taking a new model seriously — in this order:

1. Find the benchmarks

First stop, three leaderboards:

  • MMLU-Pro (general reasoning)
  • HumanEval+ / LiveCodeBench (code)
  • IFEval (instruction following)

If none of those have numbers, knock 30% off your credulity.

2. Find the inference cost

The biggest pitch for non-Transformer models is usually long-context inference cost. Three questions:

  • At 100K token context, what's the tokens/sec throughput?
  • How much VRAM does the KV cache eat?
  • Deployment floor — what model size fits on a single A100 / H100?

3. Run your own tests

Don't just trust official numbers. I usually run:

# Simple long-context retrieval test
import jev  # hypothetical package name
model = jev.load("jev-base")
context = "..." * 50000  # 50K tokens
question = "At token 47,382, which city does the author mention?"
response = model.generate(
    context=context,
    prompt=question,
    max_tokens=100,
)
print(response)
# Compare against Llama 3.1 70B on the same input

If Jev really has a "new shape" to sell, long-context retrieval should be its showcase.

4. Find the code or paper

  • GitHub repo? Star count doesn't matter — watch commit frequency and issue response.
  • arXiv paper? Check whether it has architecture diagrams but no experiments.
  • Demo URL? Can you sign up and play immediately?

5. When to trust, when to wait, when to ignore

  • Trust: open weights + third-party benchmark reproduction matches + inference cost below equivalent Llama tier
  • Wait: demo only, closed source, all benchmarks run in-house
  • Ignore: no demo, benchmarks that are just paraphrases of someone else's numbers

By that standard, Jev is squarely in the "wait and see" bucket right now.

Editorial Reflection

If "new shape" turns out to be real, the biggest narrative in the 2026 LLM world could shift from "how do we scale parameters" to "can the architecture actually change." That matters more than the next bigger-parameter GPT-5 successor — because it touches the entire inference cost curve and hardware demand structure, not just single-point performance.

But I'm cautiously optimistic about this news, for three reasons:

First, the death valley for new-architecture LLMs is ecosystem, not technology. Mamba and RWKV both died at the Hugging Face integration step. Jev either grows its own ecosystem or gets quietly absorbed into a major framework via acquisition — otherwise it'll likely vanish into the footnotes of AI history within three years.

Second, Simon's blog post isn't an endorsement. He wrote "this is worth noticing," not "this is great." He uses similar tone for Llama 4, GPT-5, and the new Claude version. Readers need to separate "noticing" from "recommending."

Third, public information is still too thin. The 20-points-3-comments data point itself shows the community hasn't converged on anything. My prediction: Jev either drops a full paper + repo within two weeks, or fades off the HN front page and gets buried by the next new model.

I give the first outcome 30%, the second 60%, and the remaining 10% to a quiet acquisition by a major lab, after which no independent news ever surfaces.

Regardless of how it ends, I'll be watching this one for the rest of 2026.

Source (must cite)

https://simonwillison.net/2026/Sep/21/jev/

Source: https://simonwillison.net/2026/Sep/21/jev/