Open-Source AI

Today's Stories

1. 1

   ELYZA free-releases ELYZA-Thinking-1.0-llm-jp-4-33b models

   KDDI Group's ELYZA released ELYZA-Thinking-1.0-llm-jp-4-33b and ELYZA-Thinking-1.0-llm-jp-4-32b-a3b for free on October 2 via Hugging Face, built on NII's LLM-jp-4 series. ELYZA said the two models scored above NII's improved LLM-jp-4.1 on some benchmarks, suggesting outside customization can lift a base Japanese model's results.

   The models' practical value hinges on which benchmarks and tasks they were measured against, since only 'some benchmarks' are cited. The same release also launched ELYZA RSI Research, whose recursive self-improvement goal is still described as pre-stage.

2. 2

   Ai2's Olmo-core 3 hits 52,000 tokens per second, topping Megatron-core

   The Allen Institute for AI announced Olmo-core 3 on Thursday, a training framework it says scaled mixture-of-experts models to over one trillion parameters while processing 52,000 tokens per second on Nvidia B3000 GPUs, about 2.7 times Nvidia's Megatron-core. The framework is meant to lower the memory and networking costs that keep trillion-parameter models out of reach for researchers without state or enterprise infrastructure, which may widen who can train these systems.

   The throughput figure is a first-party benchmark, and the test is whether outside developers reproduce it on other hardware. The project and related systems are already on GitHub.

3. 3

   Trillium Labs to open up risky AI research, aims for $40 to $100 million

   Nathan Lambert and Tom Zick launched Trillium Labs, which will publish experiment details in areas like recursive self-improvement and agents. It raised an undisclosed sum and aims to raise $40 to $100 million. Publishing how models are tuned could let outside scientists scrutinize and replicate work that frontier labs keep closed, which Lambert argues is key to mitigating risks.

   The $30 million planned for training over 18 months hinges on whether the remaining $40 to $100 million is raised. Watch whether outside researchers can replicate the published runs.

4. 4

   NetApp and Iterate.ai push AIPod Mini for private AI

   NetApp and Iterate.ai are packaging the AIPod Mini with Iterate's Generate platform and an embedded LLM, so enterprises can run AI queries locally on their own hardware. The companies say open-weight and closed models are getting close to the frontier, which the partners see as making privately run AI a credible option for enterprises that want to keep data under their own control.

   Iterate.ai's results — an insurance accident-report analysis cut from about eight hours to eight minutes, and a forensic agent that identified $17.4 million in denied claims — are early examples; the test is whether such outcomes carry over to other customers.

5. 5

   Qwen3.8-Flash-Next Previews Qwen4.x Architecture

   Qwen released Qwen3.8-Flash-Next on August 27, 2026, calling it a preview of the architecture planned for Qwen4.x. The design suggests Qwen's team is applying its own improvements rather than reusing existing structures, which points to independently developed technology rather than reliance on distillation.

   The outcome hinges on whether llama.cpp adds full support for the new attention structure before Qwen4.x arrives. Watch the sparse flash attention pull request, #28349.

6. 6

   Clef-flash vs Jev: format matches, speed claim fails

   A developer testing Cloudflare's Clef and Clef-flash, released October 1 under Apache 2.0, found Jev's official Python SDK ran on Clef after only changing the endpoint and model name, with identical answer structure. The two families of decision models look interchangeable on the code level, so teams already using Jev can compare them on their own data rather than rewriting an application to try Clef.

   Whether the speed case holds depends on region and network, since the official 38.8 ms figure is a Cloudflare-internal median; the test also found Jev stays the cheapest option per judgment.