v5-chessmaster, 516.8 MiB)Compact 516.8 MiB (Q4_K_M) Round-6 (v5-chessmaster) Surgical Mid+Upper Cortex LoRA (Layers 8..21, 0.6337% parameters) + Zero-Gradient Roof Calibration (Layers 22..23) + Exact-Leaf Branch-and-Bound CUDA Serving (serving_v6_chessmaster/) — built on AIHeikki/brainsurged-open-jev-qwen3.5-0.8b (v4).
Qwen/Qwen3.5-0.8B → AIHeikki/brainsurged-open-jev-qwen3.5-0.8b (v4, Apache-2.0)main): brainsurged-chessmaster-qwen3.5-0.8b-Q4_K_M.gguf (516.80 MiB / 541,903,872 bytes, SHA-256: 538510f6fcd57f7eb108841ba10a19767c6a45f4ec93114808849d31e9f2a1e8)4B (2.59 GiB) Model: AIHeikki/brainsurged-chessmaster-qwen3.5-4b (80.83% on the 120-position benchmark, won 1½–½ against Jev 1.13.0 in the live rematch with 24. Rb1#)75.83% on Isolated Positions Still Lost ½–1½ in Full GamesComparing Brainsurged Chessmaster 0.8B (516.8 MiB) and Brainsurged Chessmaster 4B (2.59 GiB) provides a concrete empirical case study of how solving isolated benchmark positions and playing out a complete multi-turn game can produce different outcomes:
Holdout-Chessmaster-120):Chessmaster 0.8B jumped from 10.83% (v4) → 75.83% (v5-chessmaster) (+65.00 percentage points), coming within 5.0 percentage points of Chessmaster 4B (80.83%).0.8B hit 100.00% (30/30) on single-turn passed-pawn conversion positions and 93.33% (28/30) on Mate-in-1 finalizers, while cutting its kamikaze/shuffle trap blunder rate from 39.17% → 4.17%.Jev 1.13.0 (serving_v6 Exact-Leaf Decoder on Cloud Run NVIDIA L4):Chessmaster 4B won and drew (1½–½): Held a 140-ply draw as Black (8 blunders vs. Jev's 14) and won by checkmate as White (24. Rb1# in 47 plies, 1 blunder vs. Jev's 4).Chessmaster 0.8B lost and drew (½–1½) at 54.8 ms p50 gateway engine (107.9 ms WAN p50):Jev White vs. 0.8B Black — 0–1 in 21 plies): After 10...d5, 0.8B left the c8 mating square unguarded along a multi-square diagonal/file ray (11. Qc8#). With only 24 layers (d=1024), 0.8B hit a geometric horizon ceiling on long-range multi-piece pins and mating rays.0.8B White vs. Jev Black — ½–½ in 116 plies): Playing White, 0.8B survived 116 plies against Jev 1.13.0, matching Jev on ≥300cp blunders (11 vs. 11) and committing fewer kamikaze checks (1 vs. 2), before drawing by fivefold repetition.4B & 0.8B vs. Jev 1.13.0)Full PGNs and ply-by-ply Stockfish 18 evaluations archived in reports/chessmaster_evaluations/live_matches/ and MORNING_REPORT.md.
| Match Game | White | Black | Result | Plies | Termination | ≥300cp Blunders (White / Black) |
Kamikaze Checks (White / Black) |
Brainsurged Engine p50 (WAN p50) |
|---|---|---|---|---|---|---|---|---|
Game 1 (4B) |
Jev 1.13.0 |
Chessmaster 4B |
½–½ |
140 |
Insufficient material | 14 / 8 |
4 / 1 |
130.4 ms (186.4 ms) |
Game 2 (4B) |
Chessmaster 4B |
Jev 1.13.0 |
1–0 |
47 |
Checkmate (24. Rb1#) |
1 / 4 |
0 / 1 |
202.2 ms (262.1 ms) |
Game 3 (0.8B) |
Jev 1.13.0 |
Chessmaster 0.8B |
1–0 |
21 |
Checkmate (11. Qc8#) |
2 / 3 |
0 / 0 |
54.8 ms (109.7 ms) |
Game 4 (0.8B) |
Chessmaster 0.8B |
Jev 1.13.0 |
½–½ |
116 |
Fivefold repetition | 11 / 11 |
1 / 2 |
54.8 ms (107.9 ms) |
Holdout-Chessmaster-120 & Holdout-240)Source record: exp11_chessmaster_campaign_results.json.
| Model & Decoding Stack | Holdout-Chessmaster-120 (Zero-Think) |
Holdout-Chessmaster-120 (Micro-Blink) |
Kamikaze / Shuffle Trap Blunder Rate | Poisoned Check (30) |
Queen Guard (30) |
Passed-Pawn Push (30) |
Mate-in-1 (30) |
Holdout-240 Non-Regression |
|---|---|---|---|---|---|---|---|---|
0.8B v4 + Greedy Token-1 (serving_v5) |
10.83% |
15.83% |
39.17% (47/120) |
6.67% |
16.67% |
0.00% |
20.00% |
90.41% |
0.8B v5 + Exact-Leaf Decoder (serving_v6) |
75.83% (91/120) |
73.33% |
4.17% (5/120) |
56.67% |
53.33% |
100.00% (30/30) |
93.33% (28/30) |
90.42% (100% Spend) |
4B v4 + Greedy Token-1 (serving_v5) |
34.17% |
35.00% |
22.50% (27/120) |
10.00% |
33.33% |
23.33% |
70.00% |
100.00% |
4B v5 + Exact-Leaf Decoder (serving_v6) |
80.83% (97/120) |
80.83% |
0.00% (0/120) |
66.67% (70% Blink) |
56.67% |
100.00% (30/30) |
100.00% (30/30) |
98.75% |
0.8B: v4 → v5-chessmaster)Layers 8..21):0.8B v4 BF16 weights (132 LoRA modules + 3 roof tensors).Layers 0..7) at exact v4 to preserve JSON schema compliance and customer/business feature extraction (90.42% on Holdout-240, 100.0% on Minimize-Spend).alpha=32) to Layers 8..21 (76 linear modules, 5,439,488 trainable parameters = 0.6337% of 858,418,624 total scalars) for 2 epochs (196 AdamW steps, 10.6 min on Spot NVIDIA A100-40GB, loss 0.5147 → 0.0184).torch.set_grad_enabled(False)):Layers 22..23 + lm.norm (selected_beta_temp = 0.94, 0/16 unclosed <think> loops).serving_v6_chessmaster/):_exact_bnb_search_single_stage + jev_score_chunk_logprob) + cparams.kv_unified = true (N_CTX = 32768), running at 54.8 ms p50 gateway engine latency across 35–42 legal chess moves on a single Cloud Run NVIDIA L4.llama.cpp / serving_v6_chessmaster)ollama run hf.co/AIHeikki/brainsurged-chessmaster-qwen3.5-0.8b
llama.cpp (llama-server)llama-server \
--model brainsurged-chessmaster-qwen3.5-0.8b-Q4_K_M.gguf \
--gpu-layers 99 \
--flash-attn on \
--jinja \
--reasoning-budget 0
serving_v6_chessmaster/)See serving_v6_chessmaster/README.md:
cd serving_v6_chessmaster
python3 stage_models_from_hf.py
docker build --platform linux/amd64 -t jevlab-serving-v6-chessmaster .
| Path | Artifact | Size | SHA-256 |
|---|---|---|---|
brainsurged-chessmaster-qwen3.5-0.8b-Q4_K_M.gguf |
Flagship v5-chessmaster Q4_K_M GGUF |
516.80 MiB (541,903,872 B) |
538510f6fcd57f7eb108841ba10a19767c6a45f4ec93114808849d31e9f2a1e8 |
patches_and_attestations/v5/ |
Surgical BF16 delta .safetensors & JSON attestation |
~26 MiB |
brainsurged-chessmaster-qwen3.5-0.8b-surgical-delta.safetensors |
serving_v6_chessmaster/ |
jev_jump_engine.cpp, fast_gateway.py, Dockerfile, smoke_test_build.py |
~65 KiB |
Round-6 Exact-Leaf Branch-and-Bound CUDA serving stack |
reports/chessmaster_evaluations/ |
Live PGNs, MORNING_REPORT.md, CHESSMASTER_PILOT_MATCHED_REPORT.md, JSON records |
~180 KiB |
Complete benchmark, Stockfish 18 pilot, and live match records |
Released under Apache-2.0, inheriting the license of Qwen/Qwen3.5-0.8B by the Qwen Team (Alibaba Cloud).