🎟️ Claimed my profile on the new home for #buildinpublic.
Come find me: buildinpublic.com/?ref=losti…
23
I stopped using Ollama many months ago because it doesn't offer good performance.
But, without a doubt, we should all learn how to create products that are this easy to use. They did a great job.
1
48
Elostirion retweeted
Inkling 2 Small is a large model, and it's a weird one
594B parameters, 17B active, and only ~13% on Terminal-Bench 4.0
Qwen3.8 Flash Next has 6B active parameters (125B backbone) and scores 29% on the same benchmark. In my own runs, it solved nearly 50% of the tasks with 4 attempts at xhigh.
Now the weird part: on DeepSWE v1.1, Thinking Machines reports 72.1% for Inkling 2 Small. Qwen3.8 Flash Next sits around 50%.
So Inkling 2 Small handles long repo-level software engineering far better than a much smaller model, then struggles in a terminal.
Elostirion retweeted
Replying to @bnjmn_marie
I suppose the problem resides in context length. Qwen has 256K as default, and Inkling has 1M. This is important for long running tasks, like DeepSWE ones.
I ran DeepSWE 1.1 with Qwen3.8-Flash-Next with 1M (with RoPE) , and it got 59,1%
Also this benchmarks doesn’t test all the reasoning levels.
It’s a shame, but this benchmarking companies don’t really bother to configure the model correctly before testing it.
NVIDIA is preparing a new inference service for SM120/SM121: NVIDIA/hyperflare-inference
It's still in development, but it just looks awesome.
49
Esto de confundir causa con efecto es muy humano, ¿no te parece? @nudpiedo
Voice assistants would be completely local, or bust.
There is a lot of work to do with voice assistants yet.
The emotion-identification capability that Qwen is developing is essential; furthermore, in the future, it will also be crucial for a video camera to record the user in order to detect their facial and manual gestures.
32
There is a lot of work to do with voice assistants yet.
The emotion-identification capability that Qwen is developing is essential; furthermore, in the future, it will also be crucial for a video camera to record the user in order to detect their facial and manual gestures.
When someone sighs "I'm so tired today," the right reply isn't just about the words.
You also have to hear the fatigue in their voice, then figure out if they want to vent, need comfort, or actually need help. The same "I'm fine" can mean completely different things depending on how it's said.
To improve the empathy and persona ability of Qwen-Audio-3.1-Realtime, we built a training scheme that combines semantic understanding, emotion perception, empathy strategy, and persona expression.
If this benchmarks are lawful, this model beats GLM-5.3-Flash, but being smaller and probably faster.
ERNIE 5.5 Flash Preview just landed! 🔥
Built on a new architecture designed for efficient inference, the latest open-weight foundation model in the ERNIE family delivers stronger performance in coding, 3D modeling, knowledge work and content creation.
It keeps decoding fast while posting competitive scores on Terminal-Bench, AutomationBench, and other benchmarks.
Here's the rundown:
- 280B total parameters, 14B active per token
- 1M-token context window
- Generation speed above 200 tokens/sec
Model weights will be released by November 25.
1
60
All the code benchmarks I’ve run share the same flaws:
- They use a generic harness.
- They fail to account for the model’s levels of reasoning.
- They impose strict time limits.
- They assign a score of 0 to a test that was nearly successful.
- They set context limits that are lower than what the model can actually handle.
That is why we often notice, when using these models, that they do not align with the published benchmarks.
28
Elostirion retweeted
To the massive dismay of those looking to control compute, it's way too late. CPUs and GPUs are everywhere. Models will get smarter, algorithms will get better, and bottlenecks we used to believe in will be overcome. You will lose and the Internet will remember what you did.
Elostirion retweeted
My first open-source contribution: a mixed NVFP4 + FP8 quantization of
@NousResearch's Arete-27B, following NVIDIA's own recipe.
• MLP + lm_head → NVFP4 (4-bit, block-16 scales)
• Attention → FP8 (W8A8)
• Embeddings + vision tower → BF16
• 55 GB → 20 GiB, single RTX PRO 6000
Verified against NVIDIA's reference checkpoint (1,699 quantizers, 0 diffs),
then benchmarked BF16 vs quantized side by side:
MMLU 84.1 → 83.1 (−1.0)
GSM8K 60.7 → 62.6 (+1.9)
IFEval unchanged
Quantization with measurement, not faith.
My first open-source contribution: a mixed NVFP4 + FP8 quantization of
@NousResearch's Arete-27B, following NVIDIA's own recipe.
• MLP + lm_head → NVFP4 (4-bit, block-16 scales)
• Attention → FP8 (W8A8)
• Embeddings + vision tower → BF16
• 55 GB → 20 GiB, single RTX PRO 6000
Verified against NVIDIA's reference checkpoint (1,699 quantizers, 0 diffs),
then benchmarked BF16 vs quantized side by side:
MMLU 84.1 → 83.1 (−1.0)
GSM8K 60.7 → 62.6 (+1.9)
IFEval unchanged
Quantization with measurement, not faith.
Weights, full model card, benchmark table and reproducible pipeline:
huggingface.co/veladan/Arete…
Apache-2.0. Base model by @NousResearch. vLLM ≥ 0.29 / TensorRT-LLM,
Blackwell GPUs.
39















