DCode retweeted
나 진심 저런 사람처음봄 버스에서 음란행위를햇다가 교통사고고의로 냇다가 오버워치 주작여혐을햇다가 책을 냇다가 설탕빼고 풋고추 곁들여먹다가 수능국어로 인증을 하겟다햇다가 노쇼햇다가 7등급은 자살하라고 여자괴롭혓다가 알고보니 지가 7등급이라 죽엇다고? 뭐야??
DCode retweeted
사망한 이모씨가 능력주의 신화에 거역한 강상죄로 사형당했다는 표현에 동의하면서…(물론 그는 죽음으로 도망치지 말았어야 했다)
별개로 보도윤리 내다버린 언론이나 미디어 스포츠 소비하는 듯한 시선들을 보다보면 뭐랄까, 중세에는 공개처형식이 축제였다는 말이 멀게 느껴지지 않아서 좀 섬뜩함
DCode retweeted
서울 새끼들은 한시간에 네 번 오는 경의중앙선이 배차간격이 길다고 칭얼거림.ㅋㅋㅋ
전라북도엔 하루에 네 번 오는 버스가 유일한 대중교통인 곳이 적지 않을텐데 말임.
Replying to @Qkind
그나마 있는 버스도 없어지고 자차타고 10분만 나가면 사람 없어서 다 처망한 상권 뿐인데
우리동네에서 가장유명한 랜드마크 : 망해서 폐허된 폐건물 ㅋㅋ
DCode retweeted
대법원이 김건희 여사가 정치 브로커 명태균씨로부터 무상 여론조사를 제공받은 것은 불법 정치자금을 받은 것으로 볼 수 없다고 판단했다.
khan.co.kr/article/202611191…
DCode retweeted
Was this the ultimate act of heroism, or the most heartbreaking and expensive suicide mission in history? 😭💔
DCode retweeted
The Scarlet Witch never bothered hunting the Infinity Stones for three huge canon reasons.
Using the Infinity Gauntlet nearly killed Thanos and Hulk, and killed Iron Man. As Agatha Harkness revealed in WandaVision, the Scarlet Witch is capable of "spontaneous creation", rewriting matter, minds, and reality from scratch without taking any radiation damage!
When the Stones were still intact, Wanda was a hero who only wanted to protect Vision (who literally needed the Mind Stone in his head to stay alive!). Once she lost everything and read the Darkhold, Earth-616’s Infinity Stones had already been reduced to atoms by Thanos!
In Multiverse of Madness, Wanda needed a way to travel between alternate universes to reach her sons. Because Infinity Stones only work inside their own native universe (the TVA paperweight rule!), they were useless to her compared to America Chavez’s multiversal portals!
Why hunt down six cosmic stones that burn your arm off and only work in one universe when Chaos Magic lets you rewrite reality with a thought?
#Marvel #MCU
DCode retweeted
Qwen3.8-27B has reportedly been compressed from 55.6GB to just 5.9GB, while retaining 98.2% of its aggregate benchmark performance.
That’s an 89.4% reduction in model size, achieved by PrismML, the team behind Bonsai, using its second-generation compression system, Bonsai 2.
The reported numbers are worth examining:
Original Qwen3.8-27B: 55.6GB
Bonsai 2 checkpoint: 5.9GB
Storage reduction: 89.4%
Aggregate benchmark retention: 98.2%
Previous Bonsai 27B version: approximately 95% retention
Evaluation coverage: 20 benchmarks
The interesting part isn’t just that a 27B model now occupies less storage. It’s how much of its reported capability survives such aggressive compression.
1. How does Bonsai 2 achieve this?
Bonsai 2 uses ternary quantisation, a compression technique that represents model weights using only three possible numerical values:
-1
0
+1
Most familiar quantisation formats, such as FP16, INT8 and INT4, preserve a wider range of numerical values. Ternary quantisation takes compression further by reducing each weight to one of three states.
Scaling factors are then used to recover useful numerical ranges from these compact representations.
Why is this significant?
A neural network’s weights encode the numerical relationships that allow it to recognise patterns, follow instructions, reason and generate responses.
When those weights are compressed too aggressively, some of that information can be lost. The result may be a smaller model that struggles with tasks the original handled reliably.
Bonsai 2 is designed to minimise that loss while dramatically reducing the storage required for the model.
2. What improved over the previous Bonsai version?
PrismML reports that benchmark retention increased from approximately 95% with its previous Bonsai 27B version to 98.2% with Bonsai 2.
That’s an improvement of 3.2 percentage points in aggregate retention.
The team also highlights improvements in three areas:
Coding agents
Multimodal reasoning
Long-horizon tool use
The agentic improvements are particularly interesting.
A coding agent working on an actual project needs to do much more than generate a block of code. It has to inspect repositories, understand dependencies, modify multiple files, run tests, interpret errors and make corrections.
Some tasks require dozens of decisions and tool interactions before completion.
Maintaining that kind of behaviour after compressing a 27B model into a 5.9GB checkpoint would be particularly useful for developers running local AI agents.
However, benchmark retention alone doesn’t establish how reliably the model handles every extended workflow. That needs to be tested separately.
3. Why this matters for local AI
A 5.9GB checkpoint is considerably easier to distribute, download and store than a 55.6GB version.
It could make Qwen3.8-27B more accessible to developers working with consumer GPUs, laptops and systems with limited storage.
But there’s an important distinction: checkpoint size is not the same as total inference memory.
Running the model still requires memory for:
KV cache: Stores information from previously processed tokens.
Activations: Holds intermediate values generated during computation.
Runtime buffers: Temporary memory required by the inference engine.
Other allocations: Additional overhead from the serving environment.
The actual memory requirement depends on context length, batch size, concurrency, inference engine and hardware.
So, a 5.9GB checkpoint doesn’t automatically mean the model can run comfortably with only 5.9GB of available memory.
4. Does 98.2% retention mean the model is almost identical?
Not necessarily.
The 98.2% figure is PrismML’s reported aggregate result across 20 benchmarks. It doesn’t mean every individual capability retains exactly 98.2% of the original model’s performance.
Some tasks could experience larger performance losses than others, even when the overall average remains high.
Qwen3.8-27B reportedly reached 91.90 tokens/s on an M5 Pro Mac mini with 64GB of unified memory, using TensorFold and DFlash2 speculative decoding.
What makes this result interesting is that the developer tested two TensorFold releases, 1.2.1 and 1.2.2, on the same machine within a single day. Both delivered substantial speed improvements across chat, long-context and coding workloads.
No new model weights or hardware upgrade was required
The hardware and inference setup
Machine: Mac mini with M5 Pro
Memory: 64GB unified memory
Target model: Qwen3.8-27B
Inference engine: TensorFold
Speculative decoding: DFlash2
With standard decoding, Qwen3.8-27B generated approximately 17 tokens/s
Enabling the DFlash2 drafter pushed throughput to around 56 tokens/s, more than three times the baseline
The coding workload produced the highest reported speed, reaching 91.90 tokens/s
How the two TensorFold releases performed
The first benchmark used TensorFold 1.2.1. Later that day, developer @ashxhart released version 1.2.2
Although the newer release notes focused on other hardware, the developer repeated the tests on the same M5 Pro Mac mini
The results stayed relatively close, suggesting that the DFlash2 acceleration was not limited to one specific release
Chat workloads
TensorFold 1.2.1: 3.26× speedup
TensorFold 1.2.2: 3.35× speedup
Long-context workload with a 31K-token prompt
TensorFold 1.2.1: 2.94× speedup
TensorFold 1.2.2: 2.77× speedup
Coding workloads
TensorFold 1.2.1: 5.4× speedup
TensorFold 1.2.2: 5.5× speedup
Peak reported generation speed: 91.90 tok/s
Peak memory consumption was approximately 28.5 GiB on both versions
The long-context results are particularly worth noting. Speculative decoding still delivered a substantial speedup with a 31K-token prompt, although the improvement was smaller than in the chat and coding tests
Why DFlash2 accelerates generation
Normally, an autoregressive model generates one token at a time. Each new token depends on the previous output, so the target model must repeatedly perform decoding work
DFlash2 introduces a speculative decoding process
A smaller drafter predicts several upcoming tokens. Qwen3.8-27B then verifies those predictions. When multiple proposed tokens are accepted, the target model can advance several tokens in one verification cycle
The speedup depends heavily on the drafter’s accuracy. If most proposed tokens are accepted, the system avoids many sequential decoding steps. If too many are rejected, the benefit decreases
There is another interesting detail from this test
With randomness disabled, the developer reported that the DFlash2-generated answer was identical, word for word, to the answer produced through standard decoding
That is useful evidence that the acceleration preserved the observed output in this particular test. It should not be interpreted as a guarantee that every speculative decoding setup will always produce identical results
What’s new in TensorFold 1.2.2?
Beyond the generation-speed results, version 1.2.2 addresses startup latency
Previously, the first request could be delayed while the drafter model loaded. The newer release removes that initial wait, allowing the first interaction to begin sooner
This is an important part of local inference performance. A model can generate quickly once running but still feel slow if users have to wait for it to initialise
What this means for local inference
Qwen3.8-27B did not need a different checkpoint or a hardware upgrade to achieve these reported gains
The improvement came from changing how the existing model performs generation. TensorFold’s DFlash2 integration reduced sequential decoding work and improved throughput across several workloads
The results are developer-reported measurements from one M5 Pro Mac mini, so further testing across other Apple Silicon configurations, context lengths and inference workloads would help establish how broadly the gains apply








































