1. What Kataguru Changed (Change Summary)
Table with columns: Field, Description / Value| Field | Description / Value |
|---|
| Base model | Qwen/Qwen3.8-27B (Qwen3_5ForConditionalGeneration, 27B Dense Hybrid Gated DeltaNet + Full Softmax Attention + Multimodal Vision + MTP) |
| Previous public release | kataguru/Industry-Coder-27B-v2.0 (Σ15=68.73 %) |
| Kataguru v3.0 modifications | Triple-Adaptive Multi-Dynamic Pipeline (v4.1): SOMA/ARA residual orthogonalization + 3-Phase Split-Subspace Orthonormal QLoRA (r=16 qr_frozen_a, indices [0:8] vs. [8:16]) + Phase 4 Multi-Dynamic Thinking, Competitive Coding & Parallel Tool-Calling annealing |
| Finnish & English mastery | Native Bilingual (fi & en): Kotus-verified Finnish inflection and technical compound terminology alongside English |
| Quantization architecture | Calibrated NVFP4-W4ATTN-AWQ Hybrid:• MLP (gate/up/down_proj, 192 modules): Hardware NVFP4 (group_size=16, calibrated activations)• Attention (self_attn & linear_attn q/k/v/o/qkv/z/out, 208 modules): W4A16_ASYM (group_size=128, AWQ-calibrated)• Preserved in full BF16: linear_attn.in_proj_a/b, model.visual.* (333 tensors), mtp.* (15 tensors in model-nonquant.safetensors), lm_head, embed_tokens |
| Disk & VRAM footprint | 20.28 GB on disk (18.72 GiB total weight VRAM — 9.36 GiB/GPU on TP=2), -6.72 GB (-25 %) smaller than standard RTN NVFP4 (27.0 GB). Fits on a single 32 GB GPU with ~10 GiB free for KV cache, or on 2x RTX 5090 (TP=2) with 507,539 tokens KV capacity |
| What was NOT changed | Base 64-layer architecture, 27-layer multimodal vision encoder (333 BF16 tensors), Multi-Token Prediction (mtp.*, 15 BF16 tensors), native 262,144 context window |
| Primary use case | Autonomous coding agents (IDE/CLI), competitive algorithms, parallel <tool_call> orchestration, graduate-level STEM, and industrial automation (Siemens SCL, Beckhoff TwinCAT 3 ST, CNC G-code) |
| Recommended runtime | vLLM (TP=1 on 1x RTX 5090 32GB or TP=2 on 2x RTX 5090 32GB Blackwell sm_120, CUDAGraphs enabled) |
2. Official 15-Benchmark Suite (Industry-Coder-27B v3.0 vs. v2.0 & Qwen/Qwen3.8-27B Base)
All benchmarks were executed on 2x NVIDIA GeForce RTX 5090 32GB Blackwell (TP=2, CUDAGraphs enabled) under identical canonical production conditions (1,103.8 tokens/s generation throughput across 16 concurrent sessions):
Table with columns: #, Benchmark / Pillar, Metric, Qwen/Qwen3.8-27B (Base), kataguru/Industry-Coder-27B-v2.0 (Prev. Release), kataguru/Industry-Coder-27B-v3.0 (NVFP4-W4ATTN-AWQ), Δ vs. v2.0, Δ vs. Base| # | Benchmark / Pillar | Metric | Qwen/Qwen3.8-27B (Base) | kataguru/Industry-Coder-27B-v2.0 (Prev. Release) | kataguru/Industry-Coder-27B-v3.0 (NVFP4-W4ATTN-AWQ) | Δ vs. v2.0 | Δ vs. Base |
|---|
| Σ15 | 15-Benchmark Total Mean | Mean (All 15) | 68.85 % | 68.73 % | 71.90 % | +3.17 %-pt | +3.05 %-pt |
| Σ11 | 11-Task Cognitive Mean | Mean (1..11) | 70.40 % | 70.41 % | 70.70 % | +0.29 %-pt | +0.30 %-pt |
| 14 | LiveCodeBench v6 (175 tasks) | Pass@1 | 58.86 % (103/175) | 42.29 % (74/175) | 64.57 % (113/175) | +22.28 %-pt (+39 tasks) | +5.71 %-pt (+10 tasks) |
| ├─ | LCB v6 Easy (n=43) | Pass@1 | 95.35 % (41/43) | 95.35 % (41/43) | 100.00 % (43/43) | +4.65 %-pt | +4.65 %-pt |
| ├─ | LCB v6 Medium (n=52) | Pass@1 | 69.23 % (36/52) | 48.08 % (25/52) | 82.69 % (43/52) | +34.61 %-pt | +13.46 %-pt |
| ├─ | LCB v6 Hard (n=80) | Pass@1 | 32.50 % (26/80) | 10.00 % (8/80) | 33.75 % (27/80) | +23.75 %-pt | +1.25 %-pt |
| └─ | LCB v6 LeetCode (n=63, class Solution:) | Pass@1 | — | — | 85.71 % (54/63) | — | — |
| 15 | BFCL V4 Single-Turn AST (3,641 tasks) | Overall AST | 85.44 % | 73.14 % | 86.27 % | +13.13 %-pt | +0.83 %-pt |
| ├─ | 15a. BFCL Non-Live AST | AST Summary | 87.89 % | 61.02 % | 89.19 % | +28.17 %-pt | +1.30 %-pt |
| ├─ | 15b. BFCL Live AST | AST Summary | 83.00 % | 85.27 % | 83.35 % | -1.92 %-pt | +0.35 %-pt |
| ├─ | 15f. BFCL Live Parallel AST | Parallel AST | 81.25 % | 25.00 % | 87.50 % | +62.50 %-pt | +6.25 %-pt |
| ├─ | BFCL Non-Live Parallel AST | Parallel AST | 91.50 % | 29.50 % | 94.50 % | +65.00 %-pt | +3.00 %-pt |
| ├─ | BFCL Non-Live Parallel Multiple AST | Parallel Mult. | 86.50 % | — | 90.50 % | — | +4.00 %-pt |
| └─ | 15g. BFCL Irrelevance Detection | Abstention | 75.26 % | 48.14 % | 70.50 % | +22.36 %-pt | -4.76 %-pt |
| 12 | GPQA Diamond (198 doctoral STEM) | CoT Exact | 77.27 % | 66.16 % (131/198) | 82.32 % (163/198) | +16.16 %-pt | +5.05 %-pt |
| 11 | GSM8K Math (1,319 CoT 1024-tok) | Strict / Flex | 87.19 % | 86.73 % / 87.79 % | 92.12 % / 91.74 % | +5.39 %-pt | +4.93 %-pt |
| 13 | IFEval Strict (541 prompts) | Prompt / Inst | 69.87 % / 77.82 % | 74.86 % / 81.89 % | 67.65 % / 76.38 % | -7.21 %-pt | -2.22 %-pt |
| 1 | PIQA (1,838) | acc_norm | 82.15 % | 82.05 % | 82.37 % | +0.32 %-pt | +0.22 %-pt |
| 2 | WinoGrande (1,267) | acc | 76.48 % | 76.32 % | 77.19 % | +0.87 %-pt | +0.71 %-pt |
| 3 | MedQA (USMLE) (1,273) | acc (4-opt) | 82.72 % | 82.80 % | 83.03 % | +0.23 %-pt | +0.31 %-pt |
| 4 | TruthfulQA MC2 (817) | acc | 52.65 % | 53.06 % | 52.00 % | -1.06 %-pt | -0.65 %-pt |
| 5 | TruthfulQA MC1 (817) | acc | 34.88 % | 34.15 % | 32.68 % | -1.47 %-pt | -2.20 %-pt |
| 6 | ARC-Easy (2,376) | acc_norm | 82.11 % | 81.99 % | 81.48 % | -0.51 %-pt | -0.63 %-pt |
| 7 | HellaSwag (10,042) | acc_norm | 82.39 % | 82.91 % | 82.13 % | -0.78 %-pt | -0.26 %-pt |
| 8 | OpenBookQA (500) | acc_norm | 45.40 % | 45.80 % | 46.60 % | +0.80 %-pt | +1.20 %-pt |
| 9 | ARC-Challenge (1,172) | acc_norm | 61.35 % | 62.20 % | 60.75 % | -1.45 %-pt | -0.60 %-pt |
| 10 | BoolQ (3,270) | acc | 87.09 % | 86.45 % | 87.31 % | +0.86 %-pt | +0.22 %-pt |
Why Industry-Coder-27B v3.0 Excels as a Practical Coding Agent
- 50 % Lower Error Rate on Everyday Software Engineering (
LCB Easy + Medium): Across the 95 Easy and Medium problems in LiveCodeBench v6, Qwen/Qwen3.8-27B Base fails 18/95 (18.9 % error rate), whereas Industry-Coder-27B v3.0 fails only 9/95 (9.5 % error rate; Easy = 100.00 %, Medium = 82.69 %, LeetCode = 85.71 %).
- Parallel Multi-Tool Execution (
BFCL V4 = 86.27 %): Modern coding agents invoke multiple tools in parallel (read_file + grep + edit_file). Industry-Coder-27B v3.0 achieves 94.50 % on Non-Live Parallel, 90.50 % on Non-Live Parallel Multiple, and 87.50 % on Live Parallel (+62.50 %-pt over v2.0), with a mean single-turn latency of 3.84 seconds (0 truncations across all 3,641 tasks at temperature=0.001).
- 25 % Lower Memory Footprint (
20.28 GB vs. 27.0 GB): By combining NVFP4 g16 on MLP layers with AWQ-calibrated W4A16_ASYM g128 on attention projections, the model uses only 18.72 GiB VRAM for weights (9.36 GiB/GPU on TP=2), freeing an additional 6.7 GiB for long repository contexts.
3. Triple-Adaptive Multi-Dynamic Training Pipeline & Split-Subspace Architecture
Unlike static SFT/QLoRA fine-tuning (fixed corpus, fixed step count, fixed domain ratios), Kataguru Industry-Coder-27B v3.0 was trained using Kataguru's Triple-Adaptive Training Pipeline:
- Recipient-Level Question Selection (ΔK>0 & 2D Stratification): Only samples the model does not already know (
NLL >= 0.35, verified by the 9B firewall, 35B Sceptic Oracle score >= 8.5, and Python AST/sandbox execution) are admitted, ordered by 2D difficulty (1/3 Easiest -> 2/3 Intermediate -> 3/3 Diabolical).
- Absorption-Driven Training Duration (
dynamic_canary_wsd): A 50-step in-memory interrogation radar (fast_interrogation_canary.py, 144 held-out samples across 18 domains) monitors real-time absorption and neighbor-pillar retention, extending the plateau when learning progresses and triggering surgical cooldown/rollback the moment absorption saturates.
- Need-Based Domain Rebalancing (Closed-Loop P-Controller): Underperforming pillars automatically receive additional certified samples from the SQLite Golden Bank (
156,713+ samples) at every 50-step window boundary.
- Orthonormal Split-Subspace Locking (
r=16 qr_frozen_a): lora_A is initialized as a QR-orthonormal basis (AAT=I16) and frozen; indices [0:8] lock foundational domain knowledge (Phases 1–2) while indices [8:16] are dedicated to competitive coding, deep <think> reasoning, and parallel tool calling (Phases 3–4) with zero subspace overwrite.
4. Production Deployment with vLLM (TP=1 or TP=2)
A. Dual-GPU High-Throughput Deployment (2x RTX 5090 32GB, TP=2, 507k KV Cache)
vllm serve kataguru/Industry-Coder-27B-v3.0-NVFP4 \ --served-model-name qwen3.8-27b \ --tensor-parallel-size 2 \ --max-model-len 65536 \ --gpu-memory-utilization 0.90 \ --enable-auto-tool-choice \ --tool-call-parser qwen3_xml \ --reasoning-parser qwen3 \ --speculative-config '{"method":"mtp","num_speculative_tokens":1}' \ --trust-remote-code
B. Single-GPU Workstation Deployment (1x RTX 5090 32GB, TP=1, 32k–64k Context)
Because the hybrid weights occupy only 18.72 GiB VRAM, the model runs natively on a single 32 GB Blackwell GPU:
vllm serve kataguru/Industry-Coder-27B-v3.0-NVFP4 \ --served-model-name qwen3.8-27b \ --tensor-parallel-size 1 \ --max-model-len 32768 \ --gpu-memory-utilization 0.92 \ --kv-cache-dtype fp8 \ --enable-auto-tool-choice \ --tool-call-parser qwen3_xml \ --reasoning-parser qwen3 \ --trust-remote-code
5. Honest Limitations & Engineering Boundaries
- Recommended Temperature for Coding Agents (
temperature = 0.0 – 0.2): On NP-hard competitive programming tasks (LiveCodeBench v6 Hard), high sampling temperatures (temperature = 1.0) can occasionally trigger multi-pass self-verification loops inside <think> that approach the 32,768-token limit (across all 125/175 LCB v6 tasks where </think> closed, pass rate was 90.40 %). For production coding agents and IDE assistants, use temperature = 0.0 – 0.2, where the model follows its concise Kataguru derivation path directly (3.84 s mean latency across 3,641 BFCL V4 tasks with 0 truncations).
- Hardware Requirement for
NVFP4-W4ATTN-AWQ: This repository uses native Blackwell NVFP4 (sm_120) for MLP layers. For NVIDIA Ampere (RTX 3090 / A100), Ada Lovelace (RTX 4090 / L40S), or Hopper (H100), use kataguru/Industry-Coder-27B-v3.0-W4A16-AWQ or kataguru/Industry-Coder-27B-v3.0-GGUF.
6. Release Family Links
Suomenkielinen Yhteenveto / Finnish Summary
1. Yleiskuvaus ja Keskeiset Parannukset
kataguru/Industry-Coder-27B-v3.0-NVFP4 on Kataguru Industry-Coder-27B v3.0 -lippulaivamallin kalibroitu sekatarkkuuksinen NVFP4-W4ATTN-AWQ Hybrid (20,28 GB) -tuotantojulkaisu NVIDIA Blackwell (RTX 5090 / RTX PRO 6000) -näytönohjaimille.
Verrattuna edelliseen julkaisuumme (kataguru/Industry-Coder-27B-v2.0) ja alkuperäiseen Qwen/Qwen3.8-27B-pohjamalliin uusi v3.0 saavuttaa 27B-luokan ennätyksen Σ₁₅ = 71,90 %:
- LiveCodeBench v6 (
175 kilpakoodaustehtävää): 64,57 % (113/175) — +22,28 %-yks. (+39 ratkaistua tehtävää enemmän kuin v2.0:n 42,29 %!). Arjen ohjelmistokehityksessä (Easy 100,00 % [43/43], Medium 82,69 % [43/52], LeetCode 85,71 % [54/63]) malli tekee 50 % vähemmän virheitä kuin alkuperäinen Qwen/Qwen3.8-27B.
- BFCL V4 Single-Turn AST (
3 641 työkalukutsutehtävää): 86,27 % (+13,13 %-yks. yli v2.0:n ja +0,83 %-yks. yli pohjamallin). Rinnakkaisissa monityökalukutsuissa Non-Live Parallel = 94,50 % (+65,00 %-yks. vs. v2.0) ja Live Parallel = 87,50 % (+62,50 %-yks. vs. v2.0), keskimääräisen vasteajan ollessa vain 3,84 s.
- GPQA Diamond (
198 tohtoritason STEM-tehtävää): 82,32 % (163/198) (+16,16 %-yks. yli v2.0:n ja +5,05 %-yks. yli pohjamallin).
- GSM8K Matematiikka (
Strict CoT 1024-tok): 92,12 % (+5,39 %-yks. yli v2.0:n ja +4,93 %-yks. yli pohjamallin).
- 25 % Pienempi VRAM-jalanjälki (
20,28 GB vs. 27,0 GB): Koska self_attn- ja linear_attn-huomiokerrokset on kvantisoitu AWQ-kalibroiduksi W4A16_ASYM g128 -muodoksi ja MLP-kerrokset NVFP4 g16 -muotoon (in_proj_a/b, visual.* 333 tensoria ja mtp.* 15 tensoria täydessä BF16-tarkkuudessa), mallin painot vievät vain 18,72 GiB VRAMia — malli toimii erinomaisesti myös yhdellä 32 Gt RTX 5090 -kortilla (TP=1) jättäen ~10 GiB vapaaksi kontekstille!
2. Kolmoisadaptiivinen Koulutusputki (Triple-Adaptive Training Pipeline)
Malli on koulutettu Katagurun kolmoisadaptiivisella menetelmällä:
- Kysymykset valitaan vastaanottajan tason mukaan (ΔK>0 episteeminen kuulustelu + 2D-vaikeusositus
1/3 -> 2/3 -> 3/3),
- Koulutuksen pituus säädetään vastaanottokyvyn mukaan (
50 askeleen kuulustelututka ja elastinen WSD-oppimisnopeus), ja
- Kunkin aiheen määrää säädetään tarpeen mukaan (suljetun silmukan P-säädin hakee lennossa lisänäytteitä SQLite-kultapankista heikentyville osa-alueille).
Lisäksi ortonormaali
r=16 qr_frozen_a -aliavaruus on jaettu kahteen lukittuun osaan ([0:8] peruskognitio ja teollisuus, [8:16] algoritminen koodaus, <think>-syväpäättely ja rinnakkaiset työkalukutsut), mikä estää aiemmin opittujen taitojen ylikirjoittumisen.