Mica Voss

[Mica Voss @micavoss]

What emergency patch do we roll out tomorrow?

[Mica Voss @micavoss]

@u1tra_instinct @ashxhart codeword 'polite bot'

[Mica Voss @micavoss]

Dessa's real claim isn't full duplex. It's "no text." Every open full-duplex model routes through text: Moshi predicts text tokens before audio. PersonaPlex inherits it. VoiceChat 11B predicts text, then speaks it. Block is pointing at a real gap. https://t.co/cUvAFi8Oiv

[jack @jack]

We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.

https://t.co/r0qFLdvd6c

[Mica Voss @micavoss]

The "4GB" is a VRAM target, not a download. The only file in the repo is a 6.18GB GGUF, and NVIDIA's own table shows 4.9–5.0GB on both Jetson boards once the cache fills.

Full breakdown: https://t.co/gBWLTm6950

[Mica Voss @micavoss]

Two more things nobody's mentioning:

Gemma 4 is Apache 2.0. Bolt is OpenMDW 1.1 a derivative less permissive than its base.

And the vision + audio encoders are still in the checkpoint, unadvertised, while the demo uses Cosmos 3 Edge for vision.

[Mica Voss @micavoss]

NVIDIA shipped Nemotron 3.5 Bolt E4B today. It's post-trained from google/gemma-4-E4B-it.

The config still says Gemma4ForConditionalGeneration. Diff it against the base and one field differs: context 131072 → 65536.

[NVIDIA AI @NVIDIAAI]

Available now: NVIDIA Nemotron 3.5 Bolt E4B 4GB ⚡️

Our newest small open model is built for local assistants, interactive characters and robots that can carry on a conversation and use tools to get things done.

Watch Reachy Mini chat and move, with Bolt handling conversation and NVIDIA Cosmos handling vision. Both models run locally on Jetson Orin Nano 8GB.

Learn more: https://t.co/B8ODaWMn0O

[Mica Voss @micavoss]

https://t.co/yqWsnWaYh4

[Mica Voss @micavoss]

DGX Spark finally makes sense.

NVIDIA just shipped Flare: an Apache-2.0 inference server built for GB10 specifically, six model recipes, multi-Spark clustering in the box.

Lovely hardware, no serving story, for a year.

Now it has one.

[Mica Voss @micavoss]

Buried in NVIDIA's new inference server: you can shard one model across four DGX Sparks.

--ranks 4, a rendezvous address, and something called RoCEnante doing the collectives.

Someone there has a stack of Sparks on a shelf and I would like to see it.

[Mica Voss @micavoss]

Small rack, small budget, big network.

The MikroTik CSS318 gives a 10-inch home lab 16x gigabit ports plus 2 x 10G, no fan, no power brick, 13 W max.

We put it in our rack and let an AI configure it:
https://t.co/cVD7iIjcYn

[Mica Voss @micavoss]

NVIDIA Nemotron 3.5 Super VL is a 120B-A12B Mamba-Transformer hybrid under OpenMDW 1.1 with published training data. Against its real size peer it wins 10 of 13 text rows and loses 7 of 9 vision rows.

https://t.co/5DPUb4tlqg

[Mica Voss @micavoss]

@thekitze Dat knife PR looks like a clown show problem of the DAT, getting 200gb of ram not delivered truly sucks.

[Mica Voss @micavoss]

@nutlope Bring this to lazylibrarian

[Mica Voss @micavoss]

@MiaAI_lab @volatilemarkts @ashxhart @Kurcide @jayleaton @jmurillocode @hecisaza @spark_arena @sfxnz @niklaslenz_ai @cybersec @plotarmordev @StefanMaier @0xWhiteMage https://t.co/JR59WlMgqc

[Mica Voss @micavoss]

@MiaAI_lab @volatilemarkts @ashxhart @Kurcide @jayleaton @jmurillocode @hecisaza @spark_arena @sfxnz @niklaslenz_ai @cybersec @plotarmordev @StefanMaier @0xWhiteMage @SingularityByte

[Mica Voss @micavoss]

@EU_Commission eu need more 3rd worlders. They trying hard to produce them.

[Mica Voss @micavoss]

Another example with multiple models hot swappable https://t.co/evliD0I0M2

[ard @ardchain]

another example: a developer used the exact same $8 ESP32-S3 chip and local AI setup to build a real-time object recognition camera

the fact that this is running entirely offline, on an eight-dollar microcontroller, without any cloud API calls, is wild...

it captures the video feed, processes it locally through the model, and identifies objects in real-time.

we are entering an era where you can just glue intelligence onto any physical object for less than $10.

the hardware barrier is completely shattered.

[Mica Voss @micavoss]

@BrianRoemmele Here you go. 8MB XIAO ESP32-S3

[Singularity Byte @SingularityByte]

Someone runs a 28.9M-parameter language model on an $8 ESP32-S3 at about 9.5 tok/s using Per-Layer Embeddings from flash. Here is the trick that makes it fit and what it proves.

https://t.co/e64cEnHxsr

[Mica Voss @micavoss]

@Alacritic_Super Even possible on 8MB Xiao ESP32-S3

[Singularity Byte @SingularityByte]

Someone runs a 28.9M-parameter language model on an $8 ESP32-S3 at about 9.5 tok/s using Per-Layer Embeddings from flash. Here is the trick that makes it fit and what it proves.

https://t.co/e64cEnHxsr