Neel Mishra

Applied Scientist @Microsoft | Building Graph Neural Networks, finetuning techniques & latest implementations from scratch | GenAI & deep learning optimization

[Neel Mishra @neel_epoch]

@ashokgelal @VadimYuryev Ah damn, didn’t knew that.

[Neel Mishra @neel_epoch]

@ashokgelal @VadimYuryev Sucks, advantage of living in US

[Neel Mishra @neel_epoch]

@ashokgelal @VadimYuryev True for India?

[Neel Mishra @neel_epoch]

@VadimYuryev @Apple @tim_cook India when?

[Neel Mishra @neel_epoch]

@ivanfioravanti @grok which models are being discussed here?

[Neel Mishra @neel_epoch]

@ksuniri @superalesha Ah men, its a thinking machine with 11 tps its torture

[Neel Mishra @neel_epoch]

@xfyanz @Italianclownz @Alibaba_Qwen This speed is with mtp btw

[Neel Mishra @neel_epoch]

@stevibe The randomness could be because of token sampling strategy as well, where these tested on greedy token selection strategy?

[Neel Mishra @neel_epoch]

@Tech2Wild Benchmaxxed, borderline trash imho

[Neel Mishra @neel_epoch]

Qwen3.8-27B is here — and 27B scoring 61.7 on SWE-Bench Pro is ridiculous.

For comparison:

Qwen3.8-27B — 61.7
Gemini 3.6 Flash — 58.7
GPT-5.6 Luna — 62.7
Claude Sonnet 5 — 63.2
Grok 4.5 — 64.7

That is a tiny model, relatively speaking, sitting right next to the frontier coding stack.

27B is getting way too good. This is the one I’ve been waiting to try.

#Qwen #Qwen38 #Qwen3827B #OpenSourceAI #LocalLLM #CodingAI #AICoding #LLM #AIModels #MachineLearning

[Neel Mishra @neel_epoch]

@gaborpribek @omma_ai A field guide of odd street finds: signs, utility covers, tiny sculptures. Letting people tap one foreground object when a photo contains several would make each page feel curated instead of auto-cropped.

[Neel Mishra @neel_epoch]

@cjzafir The harness may be doing as much work as the model here. Run identical DeepResearch Bench tasks and compare RACE report quality plus FACT citation accuracy. “Unfiltered” can increase recall, but it doesn’t establish accuracy or neutrality.

[Neel Mishra @neel_epoch]

@fal Useful constraint: auto-rigging works best on humanoids with clearly defined limbs. For creatures or unusual proportions, expect some manual rig cleanup even if the generated GLB and FBX look production-ready.

[Neel Mishra @neel_epoch]

@CommerceGov The concrete piece behind the visit is CHIPS LOIs: about $2.0B across nine firms, including ~$1B planned for IBM to build a U.S. foundry for quantum-grade superconducting wafers. Still LOIs, not closed awards, so the foundry depends on definitive agreements.

[Neel Mishra @neel_epoch]

@Smallzero One nuance on “free”: the open H3 path is local 768p gen. MiniMax’s full 2K workflow still leans on hosted bits, and the official API is pay-as-you-go. Short gen time is real; best still depends on whether you need local or 2K polish.

[Neel Mishra @neel_epoch]

@thezacharyyu @belvedir_ai The sub-5-min claim is mostly SDK install and trace collection. A LoRA still wants ~200 successful examples, runs can take up to two hours, and with tracing on those prompts and completions leave the box even if inference stays local.

[Neel Mishra @neel_epoch]

@LottoLabs Quick check that saves the click: HF shows 0 bytes (README + .gitattributes only), the card labels itself a placeholder, and Qwen’s author listing has no Qwen3.8-27B yet.

[Neel Mishra @neel_epoch]

@AMDServer The rack number that sticks is 72 MI455X GPUs with ~31 TB HBM4 and 260 TB/s UALink scale-up. Because UALink is software-coherent, collective efficiency and end-to-end model throughput will matter more than peak FLOPS when sizing jobs.

[Neel Mishra @neel_epoch]

@MiniMax_AI Quick litmus for Realism People: same prompt and seed at LoRA scale 0 vs 1. fal ran that across 16 configs so you can isolate what the adapter actually changes vs base H3, especially skin and eyes.