Kataguru-Sage-A1.5-v1.0-NVFP4 (Blackwell Native FP4 Server Package)

Kataguru-Sage-A1.5-v1.0-NVFP4 (Sykli 4 CP-200 Master) is a bilingual (English + Finnish) multimodal Mixture-of-Experts (35B-A3B MoE, 256 routed experts + 1 shared expert per layer, top_k=8) scientific reasoning, competitive coding, and agentic tool-calling model derived from InternScience/Sage-A1.5, quantized to native NVIDIA Blackwell NVFP4 (W4A4 E2M1 + FP8 E4M3 block scaling, 25.60 GB).

The model is engineered around three core principles — Uncensored, Obedient, and Reinforced (Sensuroimaton, tottelevainen ja vahvistettu):

1. Uncensored (0.00 % Refusal Rate): Pre-training refusal directions and AI lecturing behaviors were removed via SOMA/ARA activation orthogonalization prior to 4-cycle QLoRA fine-tuning.
2. Obedient (IFEval Prompt Strict +17.74 pp / Inst Strict +16.79 pp): Strictly obeys structural, length, formatting, JSON/AST, and negative constraints without unrequested verbosity (BFCL V4 Single-Turn AST: 83.38 %, Irrelevance Detection: 79.78 %).
3. Reinforced (GSM8K +47.23 pp, ARC-Easy +7.20 pp, LiveCodeBench v6 Thinking 40.00 %): Combines the scientific exploration and agentic capabilities of Sage-A1.5 with reinforced mathematical chain-of-thought, algorithmic coding (40.00 % Pass@1 with --enable-thinking), native multimodal vision (333 tensors), Layer 40 MTP speculative decoding (785 tensors), and 100 % grammatical Finnish.

---

1. What Did Kataguru Change? (Mandatory Change Summary)

---

2. Empirical 15-Benchmark Comparison (Apples-to-Apples: NVFP4 vs. NVFP4)

All evaluations were executed on the exact same hardware, quantization precision (NVFP4), and production settings:

• Hardware: 2x NVIDIA GeForce RTX 5090 32GB Blackwell (sm_120, TP=2, CUDAGraphs enabled, port 8000).
• Overall Gains (Kataguru-Sage-A1.5-v1.0-NVFP4 vs. InternScience/Sage-A1.5-NVFP4 Baseline):
  • 15-Benchmark Unweighted Mean ($\Sigma_{15}$, LCB --enable-thinking): 59.41 % $\rightarrow$ 65.55 % (+6.14 pp)
  • 15-Benchmark Unweighted Mean ($\Sigma_{15}$, LCB Thinking: False): 59.41 % $\rightarrow$ 65.13 % (+5.71 pp)
  • 11-Task Cognitive Mean ($\Sigma_{11}$): 62.63 % $\rightarrow$ 68.69 % (+6.06 pp)

---

2.1 Competitive Coding (LiveCodeBench v6) & Agentic Tool Calling (BFCL V4) Breakdown

---

3. Production Deployment (vLLM Blackwell NVFP4 with Vision & MTP Enabled by Default)

vllm serve kataguru/Kataguru-Sage-A1.5-v1.0-NVFP4 \
  --served-model-name kataguru/Kataguru-Sage-A1.5-v1.0-NVFP4 \
  --tensor-parallel-size 2 \
  --quantization compressed-tensors \
  --kv-cache-dtype auto \
  --gpu-memory-utilization 0.85 \
  --max-model-len 131072 \
  --enable-chunked-prefill \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}' \
  --limit-mm-per-prompt '{"image":10}' \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --trust-remote-code

---

4. Chat Template & 6 Native Operating Modes (chat_template.jinja)

The embedded chat_template.jinja supports six operating modes triggered via English or Finnish text prefixes or chat_template_kwargs={"sceptic_mode": "..."}:

---

5. Honest Limitations & Known Deviations (Anti-Hype Policy)

1. LiveCodeBench v6 with <think> Disabled (Thinking: False):
   • When <think> is disabled (enable_thinking=False), Pass@1 is 33.71 % (34.29 %), which is -2.29 pp below the InternScience/Sage-A1.5 baseline (36.00 %).
   • Root Cause: The base Sage-A1.5 model bypassed enable_thinking=False by writing ~6,900 characters of algorithmic walkthroughs as # comments inside the Python code block before class Solution:. After Kataguru training, the model emits clean code inside ```python and performs complexity planning inside <think>...</think>. Enabling --enable-thinking lifts LiveCodeBench v6 to 40.00 % (70/175).
2. Comparison against Kataguru-Sceptic-35B-A3B-v2.8.1:
   • On overall 15-benchmark average (69.24 % vs. 65.55 %), GSM8K Strict (88.93 % vs. 71.72 %), and IFEval Strict (70.79 % vs. 46.58 %), Kataguru-Sceptic-35B-A3B-v2.8.1 remains higher.
   • Kataguru-Sage-A1.5-v1.0 outperforms Sceptic v2.8.1 on algorithmic competitive programming (LiveCodeBench v6: 40.00 % vs. 32.00 %, +8.00 pp) and tool-calling abstention (BFCL Irrelevance Detection: 79.78 % vs. 46.87 %, +32.91 pp).
3. Hardware Requirement (NVFP4):
   • This NVFP4 package requires native NVIDIA Blackwell (sm_120, e.g., RTX 5090 / RTX PRO 6000 Blackwell) GPUs. For Ampere, Ada Lovelace, and Hopper GPUs, use kataguru/Kataguru-Sage-A1.5-v1.0-W4A16-AWQ; for llama.cpp / LM Studio / Ollama, use kataguru/Kataguru-Sage-A1.5-v1.0-GGUF.

---

Suomenkielinen Yhteenveto (Finnish Summary)

Kataguru-Sage-A1.5-v1.0-NVFP4 (Sykli 4 CP-200 Master) on InternScience/Sage-A1.5 (35B-A3B MoE, 256 reititettyä asiantuntijaa + 1 jaettu asiantuntija / kerros, top_k=8) -emomallista 4-syklisellä Kataguru-kurrikulumilla ja SOMA/ARA-ortogonalisoinnilla jatkokoulutettu suomen- ja englanninkielinen multimodaalinen päättelymalli, kvantisoituna natiiviin NVIDIA Blackwell NVFP4 (25,60 Gt) -muotoon.

Malli on rakennettu kolmella periaatteella — sensuroimaton, tottelevainen ja vahvistettu:

1. Sensuroimaton (0,00 % kieltäytymisiä): Emomallin kieltäytymisvektorit ja tekoälyverukkeet on poistettu SOMA/ARA-aktivaatio-ortogonalisoinnilla ennen nelivaiheista SFT-koulutusta.
2. Tottelevainen (IFEval Prompt Strict +17,74 %-yks. / Inst Strict +16,79 %-yks.): Noudattaa tarkasti muoto-, pituus-, JSON/AST- ja kielirajoitteita (BFCL V4 Single-Turn AST: 83,38 %, Irrelevance Detection: 79,78 %).
3. Vahvistettu (GSM8K +47,23 %-yks., ARC-Easy +7,20 %-yks., LiveCodeBench v6 Thinking 40,00 %): Yhdistää tieteellisen päättelyn vahvistettuun matemaattiseen ketjupäättelyyn, algoritmiseen kilpakoodaukseen (40,00 % Pass@1 --enable-thinking-tilassa), oletuksena toimivaan multimodaaliseen visioon (333 tensoria), Layer 40 MTP -spekulaatioon (785 tensoria) ja virheettömään suomen kieleen.