Accelerating AI by Learning Optimal Datatypes
Canada
Joined October 2024
- Tweets40
- Following7
- Followers180
- Likes44
ByteShape is at #EMNLP2026 in Budapest!
Come see our Industry Track poster, "Displacement Is Not Direction", in Poster Session D!
📍 Hall H, HUNGEXPO
📅 Mon, October 26, 11:00–12:30
Can't make it? Find us anytime October 25–29!
📄 arxiv.org/abs/2606.19558
#NLProc #ByteShape
ByteShape is a Startup Sponsor at #PyTorchCon 2026!
Come visit us at Booth #52 for fun interactive demos, conversations about efficient AI, and a special surprise!
📸 Take a photo and bring home a printed keepsake!
📍 San Jose Convention Center
📅 October 20–21
See you there!
ByteShape retweeted
Trying out a handful of Qwen 3.8 27B quants tonight on my 4090 (custom lovelace runtime). Tests included HumanEval+ and a custom suite from my local harness sessions.
@ByteShape came out on top. Well done!🎉
Any other quants I should check out?
ByteShape retweeted
In the past week or so alone, ByteShape’s optimized AI models were downloaded more than one million times.
The pace is clearly accelerating.
ALT ByteShape Is Taking Off banner
2/3 For Thinking benchmarks, we use medium thinking effort with the recommended sampling parameters.
Bonsai 2 reaches ~91.5% on our composite benchmark and shows a strong quality/throughput trade-off.
The goal is simply to provide a common reference point across Qwen3.8 quants
3/3 Full results: byteshape.com/blogs/Qwen3.8-…
We released ShapeLearn-Lite days after Qwen 3.8 27B, and it already delivered some of the best quality/speed trade-offs we tested.
Now full ShapeLearn is even better: 99.63% of BF16’s aggregate benchmark score at just 3.84 bpw.
Models & results: byteshape.com/blogs/Qwen3.8-…
We recently released two optimized image-generation models: Qwen-Image-2512 and Krea-2-Turbo.
We pushed both into the ~3 bpw range while preserving their ability to generate high-quality images.
Each model comes in two flavors:
(1/2)
- GGUF: broadly compatible and easy to run across different hardware
- vLLM-Omni: around 2× faster, using our optimized backend on NVIDIA GPUs
Qwen-Image-2512:
byteshape.com/blogs/Qwen-Ima…
Krea-2-Turbo:
byteshape.com/blogs/Krea-2-T…
ByteShape retweeted
These @ByteShape models likely don't get enough attention (they don't seem to be loud like many others!). I've been running for a couple weeks, so well tested. Worth trying if you run Qwen3.6. Quite a lot faster for me, on my setup as well.
huggingface.co/byteshape
We released Qwen 3.6 35B A3B GGUF quants in both NTP and MTP.
The benchmark results made one thing clear: size, speed, and quality do not move in a straight line.
GPU-5 was hard to beat. If it fits, try it first.
Blog: byteshape.com/blogs/Qwen3.6-…
We recently released our Qwen 3.5 35B A3B quants.
If your setup can run GPU-7, you should try it.
If not, we’ve got options across all hardware.
Pi → 5090.
Blog: byteshape.com/blogs/Qwen3.5-…
ByteShape retweeted
+ Hermes Agent
+ Qwen3.5 35B A3B
+ 4x parallel agents with 262k context window (each)
+ Over 200 t/s token generation + 3000 t/s prefill
+ 23.2GB total VRAM consumption on RTX 5090
It can take 5 parallel agents, 4 was the sweet spot with 2x in completion time vs 1.74x.
Dream inference.
@NousResearch
@ByteShape
@ByteShape Qwen3.5-35B-A3B-IQ4_XS-4.12bpw.gguf is lethal.
It was the most efficient Qwen3.5 35B-A3B, consuming only 23GB VRAM at FULL native context, on a RTX 5090 with -b and -ub 256.
Just shy of 20 noisy points from UD-Q6_K_XL and a minute slower. However, Unsloth's was running at 131k.
This is a custom Rust/Next.JS bench. It doesn't reflect all other categories and use-cases.
Run your own local AI coding agent
We just published a beginner guide for using @opencode with local models (@lmstudio, llama.cpp, @ollama).
Mac, Linux, WSL2, full setup + API + config.
byteshape.com/blogs/tutorial…
From “I have a model” → “I have a working coding agent”
ALT OpenCode Tutorial
GPUs are consistent. CPUs are not.
With our ByteShape Qwen 3.5 9B quants, the same models perform well across GPUs, but CPUs each have their own “favorites”.
No one-size-fits-all. Optimize for your hardware.
byteshape.com/blogs/Qwen3.5-…
ByteShape retweeted
ByteShape was quietly launched just before the year end. Two weeks ago, we announced our investment in the company. Since its launch, and with minimal fanfare on purpose, @ByteShape cumulative downloads have easily blown past 100,000. No small feat for a new startup!
ByteShape retweeted
Announcing @twosmallfishvc's investment in @ByteShape.
In short, ByteShape is delivering step-function gains in AI efficiency, including up to 7x faster training, up to 10x faster inference, plus up to 40% compression to reduce model size.
ByteShape retweeted
Excellent quantisation on Devstral Small 2
We released ShapeLearn-optimized GGUFs for:
• Devstral Small 2 24B, tuned for RTX 40/50 GPUs
• Qwen3 Coder 30B, runs everywhere, yes even the Pi
Maximum quality. Fastest TPS. Minimal compromise.
GGUFs + interactive plots are live: byteshape.com/blogs/Devstral…
ALT Every Hardware Deserves a Coder: Devstral Small 2 24B & Qwen3 Coder 30B