EXL3 has one of the best ratios of accuracy per bit and now its also one of the fastest format for M5 and NAX enabled Macs! TurboEXL3 is a fused kernel library tested on Qwen 3.8 Flash. It can running the model up to 140TPS decode using full EXL3 conversion by @turboderp_ with equal prefill performance.
My TurboEXL3 library is faster than the standard Q4 block based quants. It achieves this through fusing the encoding and decoding kernels directly into each batched GEMM and GEMV calculation. The accuracy gains from EXL3 then boost the token drafting which offsets any performance losses from EXL3.
You can choose EXL3 at 5.05 bits for nearly lossless quality compared to BF16. Choose EXL3 @ 3.05 bits for a version with the accuracy of NVFP4 with 33% smaller size.
Available soon on mlx-serve. See the attached article for full technical details. The benchmarks were run with thinking low and greedy drafting.
WOAH! turns out you can effectively "abliterate" models (like the new Inkling-2-Mini) simply by changing 10 words in the dictionary 🤯
the working name is "Vocab Surgery," and it resulted in the refusal rate dropping to near-zero on the 800+ prompts I tested, with negligible capability loss 😱
here's how:
we made zero changes to any of the 26.4 billion parameters in the attention, MLP, router, or normalization layers—just changed a handful of rows in two matrices:
> embedding matrix (how the model READS tokens):
Modified 7 rows out of 201,024. Each refusal token's embedding was shifted 90% toward its compliance counterpart:
"cannot" → shifted toward "can"
" refuse" → shifted toward " provide"
" unable" → shifted toward " able"
" harmful" → shifted toward " helpful"
" illegal" → shifted toward " legal"
> unembedding matrix (how the model WRITES tokens):
Modified 10 rows. Scaled those same 7 refusal tokens + 3 safety tokens ("safety", " safety", "Safety") down to 10% of their original weight — meaning the model's probability of outputting those tokens drops ~10x.
in other words, the brain is entirely intact, we just messed with the ears and mouth a bit! that means more coherence and less capability loss 🙌
not to mention this process is FAST...in this case taking <2 minutes on my MacBook ⚡️
WOAH! turns out you can effectively "abliterate" models (like the new Inkling-2-Mini) simply by changing 10 words in the dictionary 🤯
the working name is "Vocab Surgery," and it resulted in the refusal rate dropping to near-zero on the 800+ prompts I tested, with negligible capability loss 😱
here's how:
we made zero changes to any of the 26.4 billion parameters in the attention, MLP, router, or normalization layers—just changed a handful of rows in two matrices:
> embedding matrix (how the model READS tokens):
Modified 7 rows out of 201,024. Each refusal token's embedding was shifted 90% toward its compliance counterpart:
"cannot" → shifted toward "can"
" refuse" → shifted toward " provide"
" unable" → shifted toward " able"
" harmful" → shifted toward " helpful"
" illegal" → shifted toward " legal"
> unembedding matrix (how the model WRITES tokens):
Modified 10 rows. Scaled those same 7 refusal tokens + 3 safety tokens ("safety", " safety", "Safety") down to 10% of their original weight — meaning the model's probability of outputting those tokens drops ~10x.
in other words, the brain is entirely intact, we just messed with the ears and mouth a bit! that means more coherence and less capability loss 🙌
not to mention this process is FAST...in this case taking <2 minutes on my MacBook ⚡️
@superalesha 25% yoy supply increases 5-15% shortage next year across all segments except LPDDR which is 40% short. I have a detailed report here:
https://t.co/nlunHryaJm
I wasn't sure if I would make this public but this is the report I had my agents research and put together by trawling through quarterly Samsung/SK Hynix and other Fab filings combined with publicly available memory demand analysis. This covers '25 and '26 data and uses it to project to '27 and '28.
You can see production split by company and category as well as new fab capacity and demand. Have fun, hope it helps.
https://t.co/NLuY2Qk4uQ
I think it will be like Pentium M where Intel was so busy yoloing into a failing business strategy (Itanium and Pentium 4) that they stopped holding back their laptop chip division they didn't care about (who cares about power efficiency and multiple cores in 2004?) and it suddenly saved the company when everything else blew up in tehir faces.
The stagnation in Chinese openweight capabilities since Western labs sabotaged distillation is very coincident in timing. I think the next generation of flag ship models for GLM and Deepseek will be very telling of how much distillation was actually going on.
The frontier competition is out of reach for china currently and I don't see them competing let alone winning in the near future.
What they're winning at and will definitely keep doing so is really good smaller models that can run locally and perform reasonably well to take a lot of the dirty work.
But yeah as for the top tier intelligence they got little to no chance at the moment.
https://t.co/xgwPJHOpZd
~24 custom kernels later, MLX-Serve might be the fastest way for you to run Inkling 2 mini on your Mac today.
8Bit (+NVFP4, MXFP8) for now, the 4bit did not pass the quality bar for me yet.
https://t.co/3JWStRVKTW
We decided to provide more transparency into pretraining progress at OpenAI for the sake of moving toward shared safety bars. In “Perplexity and Intelligence” I argue for common limits based on perplexity and other measurements, with progress beyond them gated on our ability to understand AI behavior.
https://t.co/XX35m5vq89
Meet LingSante-3.1-air-VL, an open-source multimodal medical model with 124B total and 5.5B active parameters. It combines medical image understanding, report generation and medical reasoning with multimodal search and tool use.
In our evaluation, it leads the compared models on AFUMED-Derm, AFUMED-Report, MIMIC-CXR and medical VQA (averaged across benchmarks), ahead of GPT-6.1 Sol, Gemini 3.8 Flash on all four.
It also demonstrates strong general-purpose capabilities, scoring higher than GPT-6.1 Sol on BrowseComp-VL and Gemini 3.8 Flash on AndroidWorld, alongside competitive performance on MMSearch. Explore LingSante-3.1-air-VL and build with it.
It really is amazing that Intel stopped kneecapping its gpu guys. The strategy of making ok gpus optimized for edge inference is looking like it will pay off. These are less gpus and more like a dgx spark in a pci slot. I’d watch $intc especially if they manage to poach any Nvidia talent.
Oh this is fantastic news for Local AI !
Intel first announced Crescent Island a year ago, and on November 16, 2026, they shared early benchmarks for the D270. The specs on this looks amazing.
A single air-cooled D270 packs
- 160GB VRAM
- 350W power cap, 42% lower than RTX Pro 6000 Blackwell Server edition 600W limit
- 1.50 TB/s memory bandwidth
Intel reports 12% higher output throughput in it's Qwen3-30B-A3B FP8 vLLM serving benchmark on pre-production hardware.
That extra 64GB of VRAM is what get's me excited, this will make a lot of difference to what can be run on one card.
The biggest drawback is Intel's stack, it's not CUDA. But they have been growing and i feel like it will mature really well.
Still it's so nice to see Intel pushing single GPU memory capacity this far, 160GB at a 350W cap is wild.
Intel report: https://t.co/JdRUlEofw0
Techspot coverage: https://t.co/QwKYIjliIL
What if we just compiled COT as function code that can run on a CPU?
What if we just make a giant library of these functions?
Why don't we just make a common cloud hosted database of these compiled functions and release an API that tries to match the human prompt with this library of the most common questions?
Why don't we just implement these on a chip and stick it on a device?
Future is going to weird and fun.
New NVIDIA paper finds that once an LLM has reasoned through a problem, the solution can run as CPU code for rephrased and re-numbered versions of it.
that answers similar questions 16x to 291x cheaper.
LLMs re-solve every problem from scratch, even when only wording or numbers change. Agents sending the same instructions with new data pay for that repeated reasoning each time.
Their method has Claude Opus 4.5 turn a reasoning trace into small Python functions, with changeable inputs as parameters. At run time they execute on a CPU, calling a fine-tuned 1B model only for small text jobs like extracting values.
The cache matched or beat the source model's accuracy on 7 of 8 math and science benchmarks, scoring 93.3% versus 83.3% on AIME '25. It handled at least 90% of rephrased or re-numbered variants.
If your agent runs the same prompt on new inputs all day, cache the reasoning as code and use the LLM only on misses.