Why
"Runs locally" hides two very different claims, and Qwen3.8 makes the gap unusually visible. The 27B model on a 16GB machine is the ordinary claim. The 2.4T-A95B model compressed from 4.9TB to 397GB — 91% smaller, still needing roughly 450GB of RAM — is the other one, and it is only "local" for someone who owns a server. The second thing worth pinning down is the phrase vendors use for the trade-off: "retains a lot of accuracy." Unsloth's own numbers show perplexity rising 2.58 → 4.49 and top-token agreement falling 78.9% → 66.3% between the largest and smallest 1-bit variants, in exchange for about 22% less disk. Whether that trade is worth taking is a measurement, not an opinion — and it is measurable on hardware I actually have.
How it works
The compression trick is narrower than the headline suggests. llama.cpp's IQ1_S spends 1.5625 bits per weight, of which 11 are index bits into a 2048-entry codebook. The Unsloth variants simply shrink that codebook — 1024, 512, then 256 entries — which drops the index to 10, 9 and 8 bits and the weight to 1.4375, 1.3125 and 1.1875 bpw (shipped as TQ2_0, TQ1_0 and Q1_0, names picked so the Hugging Face repo lists them at all). The claim that matters is that these are plain post-training quantizations: no quantization-aware training or distillation, which is exactly what makes them cheap to produce and worth verifying independently. The plan is to run the 27B model across several of these dtypes on hardware I own and reproduce the PPL/KLD/top-p curve myself rather than cite it, then check the practical rule of thumb — RAM+VRAM ≈ quant size, past which disk offloading quietly turns a 20 tok/s figure into something else entirely.
Where it lands in Jayverse
- Number: reproduce the accuracy-vs-bit-width curve on hardware jay actually owns before trusting a vendor's "retains a lot of accuracy." If Number ever runs local inference for research, measure perplexity and top-token agreement at each bit-width rather than citing the quantization vendor's numbers.
- Devnet: budget RAM+VRAM against the quant size rule of thumb before calling anything "local." A model that needs 450GB is only local for a server Devnet owns, not a claim to repeat for a smaller box, and disk offloading past that line quietly changes the real throughput number.
- Auditor/gitboard: record which dtype and measured accuracy delta backed any agent-facing decision. "Sub-2-bit, still good enough" should be a logged number tied to a specific decision, not a marketing phrase carried into the pipeline.
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| hides two different claims | 서로 다른 두 주장을 숨기고 있다 · 표현 하나가 실제로 여러 의미를 뭉뚱그릴 때. "hides two very different claims" |
| retains a lot of accuracy | 정확도를 대부분 유지한다(마케팅 표현) · 압축 기술이 성능 손실을 축소해서 말할 때. "the phrase vendors use for the trade-off" |
| worth taking | (감수할 만큼) 가치가 있는 · 트레이드오프가 수용할 만한지 판단할 때. "Whether that trade is worth taking" |
| rule of thumb | 대략적인 경험 법칙 · 정확한 계산 대신 쓰는 어림 기준. "the practical rule of thumb" |
| quietly turns X into | 어느새 X를 ~로 바꿔버리다 · 눈에 띄지 않게 상황이 나빠질 때. "quietly turns a 20 tok/s figure into something else" |
| cheap to produce | 만들기 비용이 적게 드는 · 제작 과정의 저비용을 표현할 때. "which is exactly what makes them cheap to produce" |
| picked so that | 그렇게 되도록 일부러 고른/정해진 · 의도적으로 이름이나 조건을 맞췄을 때. "names picked so the Hugging Face repo lists them at all" |
| VRAM | 그래픽 메모리(Video RAM) · GPU 메모리 용량, 로컬 구동 가능 여부를 가르는 기준. "RAM+VRAM ≈ quant size" |
| PPL | 퍼플렉시티(Perplexity), 언어모델의 예측 성능 지표 · 값이 낮을수록 다음 토큰 예측이 정확함을 뜻함. "reproduce the PPL/KLD/top-p curve myself" |
| KLD | KL 발산(Kullback-Leibler Divergence) · 압축 전후 모델 출력 분포 차이를 재는 지표. "reproduce the PPL/KLD/top-p curve myself" |
| bpw | 가중치당 비트 수(bits per weight) · 양자화 압축률을 나타내는 단위. "1.4375, 1.3125 and 1.1875 bpw" |