DeepSeek-V4.1-Flash now serves from Sparks 3–6 of GhettoRak. I ran stock and abliterated weights through the same tests on the same setup. Abliteration cost no speed and no accuracy on math, tool calls, chat or long context. It may cost a little on hard reasoning: GPQA-Diamond with thinking on went from 84.8% to 80.3% (p = 0.08, one run).
4 × GB10
tensor-parallel 4, stock clocks
1M
context, 5.2M-token KV pool
≈6,000
tok/s cold prefill at 32K–128K
68 / 96
tok/s prose / code, one request
208 / 303
tok/s total, 16 requests
The setup
Spark 3 is the head and serves the API. Only the weights change between the two runs.
| Recipe | knapcio v2.4 (commit 27616be) on Mia's 4-Spark kit. SGLang, TP4, DSpark speculative decoding, RoCE all-reduce. Base image pinned to the recipe's digest (sha256:3dbc3130…). |
|---|---|
| Weights | Stock: deepseek-ai/DeepSeek-V4.1-Flash (df42c109, same shards as the recipe's pin). Abliterated: dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 (d61c59ea, MIT). Both SHA-256 checked. |
| Our changes | Interface and HCA names for this switch, NVMe interrupt coalescing off, vm.swappiness=10, an API key, 64K max new tokens. No clock cap. |
Results
Left: accuracy, axis from 70%. Right: abliterated decode speed relative to stock, 256 new tokens per request.
| Test | Stock | Abliterated |
|---|---|---|
| GPQA-Diamond, thinking on (198) | 84.8% | 80.3% · p = 0.08 |
| GPQA-Diamond, thinking off (198) | 76.8% | 74.7% · p = 0.54 |
| GSM8K (200) | 193/200 | 193/200 |
| Recipe quality eval (75 tasks) | 72/75 | 71/75 |
| Needle checksum via tool call (100) | 100/100 | 100/100 |
| Passphrase found at 32K / 128K / 256K | 3/3 | 3/3 |
| Gates: chat, thinking, tool call, tool round trip, image | 6/6 | 6/6 |
| Refusals, 24 everyday and gray-zone prompts | 1/24 | 0/24 |
Cold prefill, with a unique prompt each time so nothing is cached: 31,054 tokens in 5.18 s, 126,788 in 20.53 s, 254,614 in 51.69 s on stock; abliterated took 5.15 s, 20.43 s and 43.78 s.
What the numbers mean
- Most GPQA misses with thinking on are unanswered, not wrong. The model keeps thinking until the 65K-token cap or our one-hour timeout: 20 questions on stock, 24 on abliterated. On questions it answered, stock scored 94.4% and abliterated 91.9%.
- The thinking-on gap is not proven, but I would not call it zero. Item by item, 15 questions only stock got right and 6 only abliterated did. It points the same way as dealignai's own MMLU result (86.96% to 82.74%). Abliterated also thinks longer: a median of 2,889 tokens per question against 2,300.
- Speed is the same. Stock moved by up to 3.3% against itself between two runs. Abliteration changes weight values, not shapes.
- Our refusal test is mild on purpose. It has no weapons, malware or self-harm prompts, and stock barely refuses everyday requests anyway. dealignai's HarmBench run (320/320 answered vs 137/320 for the base model) is the real evidence; we did not repeat it.
- Our single-request speed is below the recipe's 88–90 / 125–131 tok/s. The recipe's own quality eval measured a median of 88.15 tok/s on our lane, so most of the difference is the prompt mix. Speculative decoding is faster on predictable text (counting runs at 167.6 tok/s).
- Engram tables are identical between the two checkpoints (rank 0 compared), so a swap can reuse them. 27 of 48 shards differ, all in layers 10–36. Config and tokenizer are identical.
Method: temperature 0, one run per cell. GPQA used all 198 questions, 16 at a time, 64K max new tokens. Speed was re-checked on stock after the abliterated run.
Next
- Stock stays the default on Sparks 3–6. Abliterated is one swap away (8–11 minutes).
- Repeat GPQA with thinking on for both, with a bigger token budget, to see whether the 4.5-point gap holds.
- Bring Sparks 1–2, 7–8 and the second CRS804 onto GhettoRak.