red2awn

I have exactly the same thought. Anyone who had used seedance 2.0 a bit can tell Gemini is a bit behind, and seedance 2.1 is on the horizontal already.

LLM-as-a-judge is quite effective method to RL a model, similar to RLHF but more objective and scalable. But yes, anthropic is making it more serious than it is. Plus DeepSeek only did it for 125k requests, significantly less than the other labs, but Anthropic still listed them first to create FUD.

The "distillation attacks" are mostly using Claude as LLM-as-a-judge. They are not training on the reasoning chains in a SFT fashion.

What's the SOTA open source or weight available turn taking model these days? I tried pipecat/smart-turn-v3 and the results are not good. It only works well when you say a short sentence in a clear voice. Anything else will cause it to wait indefinitely. Closed source API models are obviously a lot better but adds network latency and the cost adds up.

Distilling from a teacher (Opus 4.5) and scaling RL more.

> i need to wash my helicopter at the helicopter wash. it is 50m away, should i walk or fly there with my helicopter.

Sonnet 4.6: Walk! Flying a helicopter 50 metres would be more trouble than it's worth — by the time you've done your pre-flight checks, spun up the rotors, lifted off, and then safely landed again, you'd have walked there and back twice. Just stroll over.

It was mentioned that the performance of DuckDB is similar to that of Polars (among others). In that case why would one choose DuckDB over Polars? The only differentiator seems to be that you do the querying with standard SQL instead of the library specific APIs.

They are heavily focused on code. Claude Code likely generates 100 of millions lines of Python a day, make the language a little bit better with $1.5M is extremely high leverage.

Implement transformers yourself (ie in Numpy). You'll never truly understand it by just watching videos.

It does NOT. Response healing only fixes JSON syntax errors, not schema differences.

See https://news.ycombinator.com/item?id=46332119

Tokens are sampled from logits using the constraints after a normal forward pass. The forward pass is the expensive part of LLM inference which isn't affected by structured output.

Cool stuff! I don't get how all the open source inference framework have this down but the big labs doesn't...

Gemini [0] is falsely advertising this:

> This capability guarantees predictable and parsable results, ensures format and type-safety, enables the programmatic detection of refusals, and simplifies prompting.

[0]: https://ai.google.dev/gemini-api/docs/structured-output?exam...

Very confused. When you enable structured output the response should adhere to the JSON schema EXACTLY, not best effort, by constraining the output via guided decoding. This is even documented in OpenRouter's structured output doc

> The model will respond with a JSON object that strictly follows your schema

Gemini is listed as a model supporting structured output, and yet its fail rate is 0.39% (Gemini 2.0 Flash)!! I get that structured output has a high performance cost but advertising it as supported when in reality it's not is a massive red flag.

Worst yet response healing only fixes JSON syntax error, not schema adherence. This is only mentioned at the end of the article which people are clearly not going to read.

WTF

Very interesting release:

* Hybrid MoE: 2-3x faster than pure MoE transformers

* 1M context length

* Trained on NVFP4

* Open Source! Pretraining, mid-training, SFT and RL dataset released (SFT HF link is 404...)

* Open model training recipe (coming soon)

Really appreciate Nvidia being the most open lab but they really should make sure all the links/data are available on day 0.

Also interesting that the model is trained in NVFP4 but the inference weights are FP8.

Opus 4.5 is the highest quality code I've seen out of LLMs, still some way to go to match programmers who care, but much better than most people. I find it enough to let it write the code and then manually polish it afterwards.

Claude Code does not support AGENTS.md, you can symlink it to CLAUDE.md to workaround it. Anthropic: pls support!

Correct, it's breaks the single prompt, single completion assumption baked into the frameworks. Conceptually it's still prompt/completion but for low latency response you have to do streaming KV cache prefill with a websocket server.

Nice work. Are you working on streaming input/output?

Why would you use an Omni model for text only workload... There is Qwen3-30B-A3B.

They had a Flash variant released alongside the original open weight release. It is also mentioned in Section 5 of the paper: https://arxiv.org/pdf/2509.17765

For the evals it's probably just trained on a lot of the benchmark adjacent datasets compared to the 235B model. Similar thing happened on other model today: https://x.com/NousResearch/status/1998536543565127968 (a 30B model trained specifically to do well in maths get near SOTA scores)

It is an in-house closed weight model for their own chat platform, mentioned in Section 5 of the original paper: https://arxiv.org/pdf/2509.17765

I've seen it in their online materials too but can't seem to find it now.

Flash is a closed weight version of https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct (it is 30B but with addtional training on top of the open weight release). They deploy the flash version on Qwen's own chat.

None of inference frameworks (vLLM/SGLang) supports the full model, let alone non-nvidia.

The previous -Flash weight is closed source. They do have weights for the original model that is slightly behind in performance https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct

This is a stack of models:

- 650M Audio Encoder

- 540M Vision Encoder

- 30B-A3B LLM

- 3B-A0.3B Audio LLM

- 80M Transformer/200M ConvNet audio token to waveform

This is a closed source weight update to their Qwen3-Omni model. They had a previous open weight release Qwen/Qwen3-Omni-30B-A3B-Instruct and a closed version Qwen3-Omni-Flash.

You basically can't use this model right now since none of the open source inference framework have the model fully implemented. It works on transformers but it's extremely slow.

No, this is used for model alignment during post-training, not part of the system prompt. Why this is in the training data such that Claude can regurgitate it is currently unclear.

Worth noting this is not only good on benchmarks, but significantly more efficient at inference https://x.com/_thomasip/status/1995489087386771851

Check the maximum number of completion tokens allowed.

Holy em-dash fest in the examples, would have thought they'd augment the training dataset to reduce this behavior.

I'd argue SSI and Thinking Machines Lab seem to that environment you are thinking about. Industry labs that focuses on research without immediate product requirement.