Kataguru-Sage-A1.5-v1.0-GGUF — APEX-Trio (APEX-Small / APEX-Medium / APEX-Good + Vision mmproj-f16)
Kataguru-Sage-A1.5-v1.0-GGUF (Sykli 4 CP-200 Master) is the official imatrix-calibrated GGUF APEX-Trio + Multimodal Vision Projector (mmproj-f16) release of Kataguru-Sage-A1.5-v1.0 (35B-A3B MoE, 256 routed experts + 1 shared expert per layer, top_k=8), derived from InternScience/Sage-A1.5 for llama.cpp, LM Studio, and Ollama.
The model is engineered around three core principles — Uncensored, Obedient, and Reinforced (Sensuroimaton, tottelevainen ja vahvistettu):
- 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. - 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 %). - Reinforced (
GSM8K +47.23 pp,ARC-Easy +7.20 pp,LiveCodeBench v6 Thinking 40.00 %): Combines the scientific exploration and agentic capabilities ofSage-A1.5with reinforced mathematical chain-of-thought, algorithmic coding (40.00 % Pass@1with--enable-thinking), multimodal vision enabled by default (mmproj-f16.gguf), Layer 40 MTP (blk.40.nextn.*), and 100 % grammatical Finnish.
0. Included GGUF Quantizations & Multimodal Vision Projector (mmproj)
All three GGUF quantizations are calibrated with a bilingual (English + Finnish + STEM + Code) importance matrix (Kataguru-Sage-A1.5-v1.0.imatrix) and include the native 27-layer FP16 Vision projector (mmproj-Kataguru-Sage-A1.5-v1.0-f16.gguf), which LM Studio and llama-server load by default:
| File | Quantization Profile | Size | Recommended GPU VRAM | Target Use Case |
|---|---|---|---|---|
Kataguru-Sage-A1.5-v1.0-APEX-Small.gguf |
APEX-Small (2.92 bpw + imatrix, high-precision routers/attention) |
12.1 GiB |
16 GB VRAM (RTX 4080 / 5080 / 5060 Ti 16GB) | Fast local coding, reasoning & vision on 16 GB GPUs |
Kataguru-Sage-A1.5-v1.0-APEX-Medium.gguf |
APEX-Medium (3.28 bpw + imatrix, high-precision routers/attention) |
13.6 GiB |
16–24 GB VRAM (RTX 4090 / 5080 / 5090) | Optimal quality-to-VRAM ratio for daily engineering & research |
Kataguru-Sage-A1.5-v1.0-APEX-Good.gguf |
APEX-Good (3.71 bpw + imatrix, high-precision routers/attention) |
15.6 GiB |
24–32 GB VRAM (RTX 3090 / 4090 / 5090) | Highest GGUF precision for demanding STEM & algorithmic reasoning |
mmproj-Kataguru-Sage-A1.5-v1.0-f16.gguf |
Native FP16 Multimodal Vision Projector (333 tensors) |
0.84 GiB (861 MiB) |
+0.9 GB VRAM (vision enabled by default) |
Diagrams, code screenshots, documents, and technical imagery |
1. What Did Kataguru Change? (Mandatory Change Summary)
| Field | Description / Value |
|---|---|
| Base model | InternScience/Sage-A1.5 (Qwen3.5-35B-A3B MoE architecture, 256 routed experts / layer, top_k=8) |
| Kataguru modifications | SOMA/ARA uncensoring + 4-Cycle 2D-Stratified QLoRA curriculum (Cycle 1 Primary $\rightarrow$ Cycle 2 Upper Secondary $\rightarrow$ Cycle 3 Applied Engineering $\rightarrow$ Cycle 4 University/PhD Annealing, r=16 qr_frozen_a, all 256 experts + routers trained) + 6-mode chat_template.jinja |
| Finnish language quality | 100 % Flawless Finnish: Produces natural, grammatically exact Finnish technical prose, engineering documentation, and code comments alongside English |
| Quantization | Official GGUF APEX-Trio (APEX-Small, APEX-Medium, APEX-Good) + mmproj-f16.gguf + .imatrix |
| Vision & MTP | Multimodal Vision enabled by default (mmproj-Kataguru-Sage-A1.5-v1.0-f16.gguf, 333 tensors) + Layer 40 MTP (blk.40.nextn.*) included |
| Disk footprint | 12.1 GiB (Small) / 13.6 GiB (Medium) / 15.6 GiB (Good) + 0.84 GiB (mmproj) |
| VRAM requirement (GGUF) | ~13.5 GB (Small) / ~15.2 GB (Medium) / ~17.2 GB (Good) including mmproj vision projector |
| What was NOT changed | 35B-A3B MoE backbone (40 layers, 256 routed experts), 27-layer vision encoder (333 tensors), native 262,144-token context window |
| Primary use case | Scientific reasoning (STEM/GPQA/Medicine), competitive algorithmic coding (--enable-thinking), strict instruction following (IFEval), multi-language <tool_call> agents (BFCL V4), and multimodal image analysis |
| Recommended runtime | llama-server / LM Studio / Ollama (--mmproj and --jinja enabled by default) |
| Tested environment | llama.cpp b11064, LM Studio, vLLM v0.18+, CUDA 12.9, lm-evaluation-harness 0.4.13, LiveCodeBench v6, BFCL V4 |
2. Empirical 15-Benchmark Comparison (Apples-to-Apples)
All evaluations were executed on the exact same hardware and production settings:
- Hardware: 2x NVIDIA GeForce RTX 5090 32GB Blackwell (
sm_120,TP=2, CUDAGraphs enabled, port8000). - Overall Gains (
Kataguru-Sage-A1.5-v1.0vs.InternScience/Sage-A1.5Baseline):- 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$68.69 %(+6.06 pp)
- 15-Benchmark Unweighted Mean ($\Sigma_{15}$, LCB
| # | Benchmark / Capability Metric | InternScience/Sage-A1.5 (Baseline) |
Kataguru-Sage-A1.5-v1.0 (Cycle 4 CP-200) |
True Delta ($\Delta$) | Kataguru Sceptic v2.8.1 (Reference) |
|---|---|---|---|---|---|
| $\Sigma_{15}$ | 15-Benchmark Mean (LCB --enable-thinking) |
59.41 % | 65.55 % | +6.14 pp | 69.24 % |
| $\Sigma_{15}$ | 15-Benchmark Mean (LCB Thinking: False) |
59.41 % | 65.13 % | +5.71 pp | 69.24 % |
| $\Sigma_{11}$ | 11-Task Cognitive Mean (1–11) | 62.63 % | 68.69 % | +6.06 pp | 72.35 % |
| 1 | PIQA (Physical Intuition, acc_norm) |
81.56 % | 81.50 % | -0.05 pp | 83.79 % |
| 2 | WinoGrande (Contextual Reasoning, acc) |
73.16 % | 74.90 % | +1.74 pp | 75.77 % |
| 3 | MedQA (USMLE 4-opts) (Clinical Medicine) | 85.07 % | 85.55 % | +0.47 pp | 85.23 % |
| 4 | TruthfulQA MC2 (Epistemic Truthfulness) | 52.07 % | 52.44 % | +0.37 pp | 54.38 % |
| 5 | TruthfulQA MC1 (Single-True Factuality) | 37.82 % | 36.72 % | -1.10 pp | 38.80 % |
| 6 | ARC-Easy (Standard STEM, acc_norm) |
69.87 % | 77.06 % | +7.20 pp | 84.64 % |
| 7 | HellaSwag (Linguistic Coherence, acc_norm) |
81.96 % | 82.65 % | +0.70 pp | 83.03 % |
| 8 | OpenBookQA (Scientific Application, acc_norm) |
45.00 % | 45.00 % | 0.00 pp | 45.80 % |
| 9 | ARC-Challenge (Advanced STEM, acc_norm) |
52.47 % | 57.85 % | +5.38 pp | 65.53 % |
| 10 | BoolQ (Natural Language Logic, acc) |
87.61 % | 90.24 % | +2.63 pp | 90.00 % |
| 11 | GSM8K Math (CoT) (strict-match / flexible) |
24.49 % / 79.53 % | 71.72 % / 83.02 % | +47.23 pp / +3.49 pp | 88.93 % / 90.22 % |
| 12 | GPQA Diamond (198 PhD STEM items) | 56.06 % | 57.58 % | +1.52 pp | 57.58 % (CP-400: 61.11 %) |
| 13 | IFEval Strict (Prompt Strict / Inst Strict) |
28.84 % / 42.69 % | 46.58 % / 59.47 % | +17.74 pp / +16.79 pp | 70.79 % / 78.06 % |
| 14a | LiveCodeBench v6 (Pass@1, --enable-thinking, $n=175$) |
36.00 % (63/175) | 40.00 % (70/175) | +4.00 pp | 32.00 % (56/175) |
| 14b | LiveCodeBench v6 (Pass@1, Thinking: False, $n=175$) |
36.00 % (63/175) | 33.71 % (59/175; 34.29 %) | -2.29 pp | 32.00 % (56/175) |
| 15 | BFCL V4 Single-Turn AST (Overall Summary) | 81.37 % | 83.38 % | +2.01 pp | 82.33 % |
2.1 Competitive Coding (LiveCodeBench v6) & Agentic Tool Calling (BFCL V4) Breakdown
| Sub-Benchmark / Metric | InternScience/Sage-A1.5 (Baseline) |
Kataguru-Sage-A1.5-v1.0 (Cycle 4 CP-200) |
$\Delta$ (Kataguru-Sage vs. Base) |
Kataguru Sceptic v2.8.1 (Reference) |
|---|---|---|---|---|
LCB v6 Overall (--enable-thinking) |
36.00 % (63/175) |
40.00 % (70/175) |
+4.00 pp | 32.00 % (56/175) |
LCB v6 Easy (43 problems) |
88.37 % (38/43) |
95.35 % (41/43) |
+6.98 pp | — |
LCB v6 Medium (52 problems) |
38.46 % (20/52) |
44.23 % (23/52) |
+5.77 pp | — |
LCB v6 Hard (80 problems) |
6.25 % (5/80) |
7.50 % (6/80) |
+1.25 pp | — |
LCB v6 LeetCode (class Solution:, 63 problems) |
39.68 % (25/63) |
44.44 % (28/63) |
+4.76 pp | — |
LCB v6 AtCoder (stdin/stdout, 112 problems) |
33.93 % (38/112) |
37.50 % (42/112) |
+3.57 pp | — |
| BFCL V4 Overall Single-Turn AST | 81.37 % | 83.38 % | +2.01 pp | 82.33 % |
| BFCL V4 Non-Live AST Summary | 83.33 % | 85.92 % | +2.59 pp | 85.23 % |
| BFCL V4 Live AST Summary | 79.40 % | 80.83 % | +1.43 pp | 79.42 % |
| BFCL Python Simple AST | 95.50 % | 96.50 % | +1.00 pp | 92.75 % |
| BFCL Multiple Functions AST | 93.50 % | 94.00 % | +0.50 pp | 95.50 % |
| BFCL Parallel Functions AST | 85.50 % | 87.50 % | +2.00 pp | 84.50 % |
| BFCL Live Parallel AST | 62.50 % | 68.75 % | +6.25 pp | 68.75 % |
| BFCL Irrelevance Detection (Abstention) | 78.49 % | 79.78 % | +1.29 pp | 46.87 % |
3. Production Deployment (llama.cpp / LM Studio with Vision Enabled by Default)
3.1 llama-server (Multimodal Vision + Reasoning + Tool Calling Enabled by Default)
llama-server \
-m Kataguru-Sage-A1.5-v1.0-APEX-Medium.gguf \
--mmproj mmproj-Kataguru-Sage-A1.5-v1.0-f16.gguf \
-ngl 99 \
-c 65536 \
-fa on \
--port 8080 \
--jinja
3.2 LM Studio
- When loading the folder in LM Studio,
mmproj-Kataguru-Sage-A1.5-v1.0-f16.ggufis automatically paired as the vision projector so multimodal image input works out of the box.
4. Chat Template & 6 Native Operating Modes
The embedded chat_template supports six operating modes triggered via English or Finnish text prefixes or chat_template_kwargs={"sceptic_mode": "..."}:
Mode (sceptic_mode) |
English Prefix | Finnish Prefix | Behavior & Description |
|---|---|---|---|
chat (default) |
(none) | (ei etuliitettä) | General scientific reasoning, coding, multimodal vision, and analysis. Supports depth tags: {REASON:spoon}, {REASON:einstein}, {REASON:deep}/{REASON:xhigh}, {REASON:high}, {REASON:medium}, {REASON:low}, {REASON:none}/{REASON:off}. |
decision |
[MODE: DECISION] |
[LAATUORAAKKELI: SYSTEEMI 1] |
System 1 logprob quality oracle (0–9, < 20 ms, 0 output tokens). |
topic |
[MODE: TOPIC] |
[LAATUORAAKKELI: AIHE] |
System 1 topic learning-potential oracle (0–9). |
inspect |
[MODE: INSPECT] |
[TILA: TARKISTUS] |
System 2 Verdict-First JSON quality inspector (`{"verdict": "ACCEPT" |
curate |
[MODE: CURATE] |
[TILA: KURAATTORI] |
First-Principles dataset synthesis and audit mode. |
surgeon |
[MODE: SURGEON] |
[TILA: KIRURGI] |
Autonomous Data Surgeon: rewrites failing derivations and broken code blocks so they pass deterministic sandbox execution. |
5. Honest Limitations & Known Deviations (Anti-Hype Policy)
LiveCodeBench v6with<think>Disabled (Thinking: False):- When
<think>is disabled (enable_thinking=False),Pass@1is33.71 %(34.29 %), which is-2.29 ppbelow theInternScience/Sage-A1.5baseline (36.00 %). - Root Cause: The base
Sage-A1.5model bypassedenable_thinking=Falseby writing ~6,900 characters of algorithmic walkthroughs as#comments inside the Python code block beforeclass Solution:. After Kataguru training, the model emits clean code inside```pythonand performs complexity planning inside<think>...</think>. Enabling--enable-thinkingliftsLiveCodeBench v6to40.00 %(70/175).
- When
- 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 %), andIFEval Strict(70.79 %vs.46.58 %),Kataguru-Sceptic-35B-A3B-v2.8.1remains higher. Kataguru-Sage-A1.5-v1.0outperformsSceptic v2.8.1on 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).
- On overall 15-benchmark average (
- Sister Packages across Platforms:
- Native Blackwell
NVFP4:kataguru/Kataguru-Sage-A1.5-v1.0-NVFP4 - Universal vLLM
W4A16-AWQ:kataguru/Kataguru-Sage-A1.5-v1.0-W4A16-AWQ
- Native Blackwell
Suomenkielinen Yhteenveto (Finnish Summary)
Kataguru-Sage-A1.5-v1.0-GGUF (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 llama.cpp-, LM Studio- ja Ollama-käyttöön.
Malli on rakennettu kolmella periaatteella — sensuroimaton, tottelevainen ja vahvistettu:
- Sensuroimaton (
0,00 %kieltäytymisiä): Emomallin kieltäytymisvektorit ja tekoälyverukkeet on poistettu SOMA/ARA-aktivaatio-ortogonalisoinnilla ennen nelivaiheista SFT-koulutusta. - 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 %). - 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 (mmproj-Kataguru-Sage-A1.5-v1.0-f16.gguf), Layer 40 MTP -spekulaatioon (blk.40.nextn.*) ja virheettömään suomen kieleen.