EvoX Genesis for breakthrough AI development and Institute of Foundation Models Launches the Industry’s Largest Fully Open-Source Fleet of AI Models

Offtopic
1

In short, it made a C compiler (C11 Compliance & C23 Subset) from scratch for US$44.38, and ported 462K-LOC Fortran 90/2008 stellar astrophysics suite to 67K LOC of pure Rust library code (13 idiomatic Rust crates) for US$10.64: Idiomatic Julia code in AI generated code - #7 by Palli

It’s clear to me the model used is no longer the main thing, rather the harness or such tools used with. That said (despite the tool above used DeepSeek and GLM), I think this model is seriously intriguing and I would look into them:

MoVA: scaling attention with sparse experts

The core idea behind mixture-of-experts is to increase total model capacity while keeping the computation required for each token roughly fixed. ..

Our new architecture, MoVA—Mixture-of-Value Attention, extends this principle to attention. .. introducing sparsity there opens another dimension for scaling model capacity beyond the feed-forward network.

MoVA integrates expert routing into multi-head attention while remaining compatible with efficient techniques including FlashAttention, grouped-query attention, and sparse attention.

The result is K2 Horizon MoVA 36B-A4B: a model with 36 billion total parameters but approximately 4 billion active parameters per token. Under the same training conditions, it performs only slightly below the dense Horizon 32B model while requiring substantially fewer active parameters.

To quantify diversity at corpus scale, we developed a new compressor, Wzip, that combines a novel adaptive multi-sliding-window LZ77 algorithm with windowed Huffman coding. Wzip mitigates the rapid saturation observed with conventional gzip/zstd compression metrics as the number of documents grows, enabling more meaningful diversity measurements across large corpora.

Horizon 36B-A4B reaches nearly the performance of the dense 32B model while activating only approximately 4 billion parameters per token. Its efficiency comes from MoVA, our new sparse attention architecture, together with MoE feed-forward layers.

These two models serve as an important reference point for studying how dense and sparse architectures behave under similar training conditions.

Across reasoning, mathematics, coding and agentic tasks, K2 Horizon delivers top-tier performance in every size class—with the 0.9B, 3.7B and 7B models setting new state of the art at their respective scales.
..
Open source is much more than open weights. .. Every model in the fleet ships with its training data, recipe and evaluations. This is open science, and we believe it’s the best path forward for AI.

K2 Horizon includes key technology innovations:

  • A unique “diffusion distillation” technique that generates blocks of tokens in parallel, which speeds model performance by roughly 3X without degrading response quality.
  • A “mixture of value attention” architecture that improves reasoning without adding more computation.
  • 36B (4B active): a new Mixture of Value Attention (MoVA) architecture that outperforms many larger models while activating only 4B parameters, which makes the model cost-effective for local hosting.
  • 375B (23B active): K2 Horizon’s flagship model, built to compete with open-weight models on reasoning and agentic work, and ready for the enterprise.

DeepSeek V4.1 is also out seemingly very good. And of course OpenAI’s Astra 6, and Anthropic’s Fable 5.1, the competition is heating up, and I read OpenAI changed to a looped model (but not something I can confirm, like also suspected for Mythos/Fable), likely one reason it’s so good (and its “computer use”). Note, for Astra on the AGI benchmark ARC-AGI-3:

62.7%, $26.1K (Standard) · 98.6%, $17.3K (Provider Adapter)

vs. 100.0% at $2,986 for Tyco at ARC Prize - Community Leaderboard showing how much the harness/adapter can matter.

2

With all due respect, I believe Qwen 3.8 Flash Next was overlooked by you. In my opinion, it is the most impactful of the recent releases. In BF16, it runs on 512GB VRAM with 650k context; additionally, you can move the entire 51B N-gram embedding table to RAM without any performance degradation. This is my current daily driver. Phenomenal performance. I use it mostly for PDF distillation together with Prime Agent. I haven’t tried EvoX Genesis yet, do you think it’s better than Prime Agent, if I may ask?

3

From a layman’s perspective, Genesis seems to be more like a construction site manager, who decides who does what and checks the quality of their work. The actual construction is done by whichever AI model is hired for the job, and most importantly, it can swap workers mid-project and the building and its history will still stand.

Genesis is some kind of Chris Rackauckas, but for AI coding agents.

4

I’ve only set up EvoX Genesis, so far, not really used it. I still think Prime Agent is very intriguing. Both are recursive, though I think it doesn’t mean the same for both.

I believe they are mutually exclusive (i.e. not to be used together, though you could start with one, then later use Prime Agent on the same codebase), so I will be looking at EvoX Genesis going forward. It’s “Not just a harness”, I recommend reading about it, including here from the developer (you can also ask):

As far as we can determine, Genesis is the first publicly known autonomous system to complete and submit a result for a Terminal-Bench Challenge. Genesis completed the WASM Render challenge for just $36, far below Terminal-Bench’s stated expectation of $1K+ per challenge.

And in some of his relies:

If you check our development history, Genesis is a direct descendant of EvoGit, which we developed about a year ago and used to win the AgentX Competition 2025.

In practice, modern coding models handle this kind of judgment reasonably well, as long as it’s not too small (7B model might have trouble doing this).

Well it made me smile, but I think not yet (at least for his expertise, differential equations). Yes, he and Genesis are both very capable. @ChrisRackauckas wrote in February that vibe coding IS for experts, but only for experts (I wander how his views might have changed, if AI assisted/autonomous coding is now for more than just experts):

I don’t know you, but I do know that you think you’re too good to vibe code. And guess what, you’re absolutely and totally wrong.
..
No, you’re not too good to vibe code. In fact, you’re the only person who should be vibe coding.

I would have thought this statement was crazy just a month ago because this label of “expert” coder also applies to me. ..

A Mental Model for LLM Agents: Your Sophomore Year Student/Intern

Drop the hype, I’m not here to sell you a ChatGPT so I’m not going to tell you it’s PhD level when it 100% absolutely clearly isn’t to anyone who has ever met a PhD in their life. But it is something, what is it? …
..

Major Point: Vibe coding turns everyone into a team lead, but not everyone should be a team lead

Since it’s not allowed to post AI answers (I suggest you try the AI questions, I could send you my answers in private):

Is EvoX Genesis a harness and an alternative or can be used together with prime-agent?

Not really, I’ve mentioned Qwen 3.8 before, the large one, and I believe the Flash version is also a big deal. I’m just also exited about K2 Horizon, mainly based on their “new architecture, MoVA—Mixture-of-Value Attention” and their claims. Feel free to use it or Qwen or any model, I think ExoX Genesis makes the exact model you choose less important. You can even use more than one model.

You can even use more then one model.

My excitement for EvoX Genesis is based on results, like same results after porting, sometimes bitexact, and up to 6.87× faster and smaller Rust code base):

Table S18: Numerical agreement and runtime performance for six migrated workloads.

Workload Fortran (s) Rust (s) Speedup Checksum difference
End-to-end burn 0.446 0.287 1.55× 3.1×10−9
EOS lookup 1.786 1.115 1.60× Bit-exact
Opacity lookup 0.532 0.269 1.98× 1.3×10−13
2D interpolation 0.159 0.101 1.58× 4.9×10−12
ROS2 integration 0.722 0.136 5.30× 5.1×10−15
Newton solve 0.084 0.012 6.87× Bit-exact

A separate timing check repeated the burn proxy 40 times per implementation after three warm-up runs, again pinned to CPUs 0–3. The median was 0.2996 s for Fortran and 0.2427 s for Rust, a 1.23× ratio. Thirty of the 40 Rust timings were below the Fortran median; this is not a 30/40 paired win rate because the paired run vectors are unavailable. Both implementations reported the same integrator counts: 8,889 function evaluations, 199 Jacobian evaluations, 199 steps, 198 accepted steps and one rejected step. The abundance-checksum relative difference was 3.1×10−9.

Table S19: Separate 40-run timing check for the burn proxy.

Statistic Fortran (s) Rust (s) Fortran/Rust
Minimum 0.2897 0.2372 1.22×
P10 0.2922 0.2385 1.22×
P25 0.2927 0.2392 1.22×
Median 0.2996 0.2427 1.23×
Mean 0.3447 0.2735 1.26×
P75 0.4416 0.3609 1.22×
P90 0.4458 0.3729 1.20×
Maximum 0.4490 0.3815 1.18×

The 1.55× burn ratio in the six-workload table and the 1.23× ratio in the separate 40-run check come from different batches and should not be pooled. The package contains summary statistics for the 40-run check but not the 80 individual timings, so we cannot reconstruct a confidence interval or paired test. Timing was also sensitive to host load: one noisy interleaved run gave a ratio of 0.82×, while a quieter rerun gave about 1.22×. We therefore use the 1.23× 40-run median as the more conservative burn result and do not claim a general speed advantage.

5

I can’t make EvoX Genesis working on my hardware, however, based on your description of its abilities, Prime Agent does similar things but I can’t comment on details. Qwen 3.8 Flash Next is a bit different then 3.8 27B or 3.8 Max, it’s a prequel to their new architecture, internally marked as 4.