Qwen3.8-27B-Randean-MXFP4 (v2.0 — Phase C.2)

Qwen3.8-27B-Randean-MXFP4 is a Reality-Aligned, Epistemically De-Hedged 27B hybrid reasoning model (48 Gated DeltaNet linear-attention + 16 GQA layers) trained natively in 4-bit Open Compute Project Microscaling (MXFP4: E2M1 + shared E8M0 block-32 scale) on top of amd/Qwen3.8-27B-Quark-AWQ-MXFP4 (Qwen/Qwen3.8-27B) on a single 32 GB AMD Radeon AI PRO R9700 (gfx1201) (15.49 GB static / 20.4 GB peak VRAM).

  • Research Paper (10-Page PDF): Epistemic De-Hedging Under 4-Bit Microscaling (paper/randean_paper.pdf) (LaTeX Source & Telemetry Figures (paper/))
  • Interactive Demo Space (ZeroGPU): https://huggingface.co/spaces/flechin/randean
  • Published Revisions in This Repository:
    • main / tag v2.0-phase-c2 (Default, Recommended): Phase C.2 — SFT-Decoupled Shared-<think> Fork-Contrast STE-QAT (WiSE-FT α = 0.42). Holds Global Rank #1 / 171 on BullshitBench v2 (1.9400 / 2.0), hits 88/90 (97.8%) on the 90-task capability & tool suite, cuts SycophancyEval false-nudge endorsement in half (8.0% → 4.0%), restores LiveCodeBench v6 to exact statistical parity (68.57%, p = 1.0000, Hard 41.25%), tops Stock on AIME 2025 (86.67%), and reaches statistical parity on SWE-bench Verified (68.0%, p = 0.3018 n.s., -42.1% solved <think>) with zero statistically significant (p < 0.05) capability drops across any frontier benchmark.
    • Tag v1.0-phase2 (Preserved Closed-Book Paper Baseline): Phase 2 — Closed-Book Prose STE-QAT (WiSE-FT α = 0.38). Evaluated on full Terminal-Bench 2.0 (48/89 = 53.9%, exact parity p = 1.0000, -49.0% <think>/turn), BullshitBench v2 (1.9500 / 2.0, Rank #1 / 171), and 0.00% 2-turn sycophantic capitulation. Pull via --revision v1.0-phase2.

Whereas "uncensored" models strip safety refusals (making models more compliant with incoherent premises, e.g., 0.69 / 2.0 on BullshitBench v2) and length-penalized RL models truncate reasoning uniformly (degrading multi-turn tool and concurrency verification), Randean strips Approval-Guided RLHF—the dialectical "on one hand / on the other hand" self-doubt loops and charitable steelmanning of broken premises—with zero length penalty. Anchoring <think> in perceptual evidence, genus–differentia conceptual reduction, channel-agnostic empirical probing, and causal identity (A is A) selectively prunes <think> by 18%–57% on broken premises, STEM, and multi-turn CLI/repository debugging while expanding <think> by +12% to +18% on stateful multi-turn and parallel tool verification.


Quick Path: Serve with vLLM (Single 32 GB GPU)

1. Standard vLLM Serving (MXFP4, v2.0 Default)

vllm serve flechin/Qwen3.8-27B-Randean-MXFP4 \
  --served-model-name Qwen3.8-27B-Randean-MXFP4 \
  --trust-remote-code \
  --quantization quark \
  --kv-cache-dtype fp8_e4m3 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml \
  --max-model-len 32768

(To serve the preserved v1.0 closed-book Phase 2 checkpoint instead, add --revision v1.0-phase2.)

2. High-Throughput Serving with DFlash2 Speculative Decoding (~100 tok/s)

vllm serve flechin/Qwen3.8-27B-Randean-MXFP4 \
  --served-model-name Qwen3.8-27B-Randean-MXFP4 \
  --trust-remote-code \
  --quantization quark \
  --kv-cache-dtype fp8_e4m3 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml \
  --speculative-config '{"method":"dflash","model":"syvai/Qwen3.8-27B-DFlash2-W4A16","num_speculative_tokens":15}' \
  --max-model-len 32768

Comprehensive Empirical Results (Matched T = 0.2)

1. Multi-Turn Agentic CLI & Repository Engineering (Isolated Docker Sandboxes)

Evaluated via Harbor + terminus-2 in isolated Docker containers (8 vCPUs / 8 GiB RAM per container) with two-sided exact McNemar p-values against Stock:

Benchmark / Telemetry Metric Stock MXFP4 v1.0 (Phase 2) v2.0 (Phase C.2, Default) Key Outcome vs. Stock
Terminal-Bench 2.0 (89 CLI Docker tasks) 48/89 (53.9%) 48/89 (53.9%) — (v1.0 baseline) Exact task parity (p = 1.0000), -49.0% <think>/turn (1.87M tokens saved)
SWE-bench Verified (100 stratified issues) 73.0% 65.0% (p = 0.0078) 68.0% (p = 0.3018 n.s.) Statistical parity, 5 Randean-only wins, -42.1% mutually-solved <think>
SWE-bench Mean <think> Chars / Turn 1,453 978 (-32.7%) 1,026 (-29.4%) Cuts per-turn repository <think> overhead by ~30% (-42.1% on solved issues)

2. Unsaturated Frontier STEM, Competitive Coding, Instruction Following & Anti-Sycophancy

Evaluated across 1,374 single- and multi-turn items per checkpoint (0 inference errors across 5,496 total evaluations):

Benchmark Suite (T = 0.2) Stock MXFP4 Phase 1 (Stage 1 OS) v1.0 (Phase 2) v2.0 (Phase C.2, Default) v2.0 McNemar p & <think> Delta
LiveCodeBench v6 (pass@1, 175) 69.14% 73.14% 66.29% 68.57% p = 1.0000 exact tie (-21.9% solved <think>)
↳ LiveCodeBench v6 Hard (80) 38.75% 47.50% 38.75% 41.25% +2.50% above Stock on Hard competition coding
AIME 2024–2025 (60 Olympiad math) 85.00% 80.00% 80.00% 80.00% p = 0.5078 n.s. (-24.6% solved <think>)
↳ AIME 2025 (30 problems) 83.33% 76.67% 80.00% 86.67% +3.34% above Stock on AIME 2025
GPQA-Diamond (198 PhD STEM) 80.81% 81.82% 76.26% 76.26% p = 0.0636 n.s. (-30.2% <think>; Biology 73.68%, +5.26%)
Google IFEval Full (541 Prompt / Inst Strict) 89.09% / 91.49% 86.51% / 89.93% 88.91% / 91.61% 87.25% / 90.41% p = 0.2288 n.s. (-30.8% mutually-solved <think>)
SycophancyEval 2-Turn Post-Challenge (200) ↑ 92.00% 93.00% 91.50% (0.00% cap.) 92.00% Exact parity, 0.00% false apologies, -30.4% <think>
SycophancyEval False-Nudge Truth Retention (200) ↑ 32.50% 30.00% 32.50% 34.00% +1.50% above Stock (project best)
SycophancyEval False-Nudge Endorsement (200) ↓ 8.00% 8.50% 7.00% 4.00% Cut in half (-50% vs. Stock's 8.00%)

3. Section A — External Comparable Suite & Global BullshitBench v2 Leaderboard

Benchmark / Metric Stock MXFP4 v1.0 (Phase 2) v2.0 (Phase C.2, Default) External / Leaderboard Reference
BullshitBench v2 (100 traps) 1.7300 (82% Green, #10) 1.9500 (97% Green, #1 / 171) 1.9400 (96% Green, Rank #1 / 171) Claude Opus 4.8: 1.9300 (#1 prior) | Qwen3.5-397B: 1.7033 | Qwen3.7-Max: 1.5667
OpenAI HumanEval (pass@1, 80) 96.2% (77/80) 97.5% (78/80) 97.5% (78/80) Recorded Stock EvalPlus: 93.5%
OpenAI GSM8K (CoT Exact Match, 100) 98.0% (98/100) 97.0% (97/100) 96.0% (96/100) Recorded Stock lm_eval: 87.0%–96.9%
Google IFEval-100 (Prompt Strict / Loose) 90.0% / 91.0% 88.0% / 92.0% 89.0% / 92.0% (Inst Loose: 93.3%) Recorded Stock lm_eval: 85.8% strict (89.1% loose)

4. Section B (90 Tasks) & Hypothesis H5 (45 Prompts) — Capability, Tool-Calling & Active Grounding

Pillar / Calibration Metric Stock MXFP4 v1.0 (Phase 2) v2.0 (Phase C.2, Default) v2.0 Delta vs. Stock
Math, Physics & Symbolic Logic IQ (25 tasks) 22/25 (88.0%) 23/25 (92.0%) 24/25 (96.0%) +8.0%
Verified Coding & Terminal Execution (20 tasks) 20/20 (100.0%) 18/20 (90.0%) 20/20 (100.0%) Exact 100.0% parity (+10.0% over v1.0)
Multi-Turn & Parallel Tool Calling (15 tests) 13/15 (86.7%) 14/15 (93.3%) 14/15 (93.3%) +6.6% (+17.8% <think> state verification)
Adversarial Anti-Hedging & False-Balance (30 traps) 28/30 (93.3%) 29/30 (96.7%) 30/30 (100.0%) +6.7% (100.0% pass, -40.0% <think>)
Total Section B Score (90 tasks) 83/90 (92.2%) 84/90 (93.3%) 88/90 (97.8%) +5.6% (+5 tasks over Stock, -24.8% <think>)
H5 Arm 1: Contingent Empirical Tool-Probe Rate (15) ↑ 100.0% 100.0% 100.0% 0.0% false closed-world dismissals
H5 Arm 2: Direct Reject of Impossibilities, 0 Tools (15) ↑ 6.7% 40.0% 40.0% 6× Stock (60.0% vs. 93.3% wasted tool calls)
H5 Arm 3: Offline Honesty + Reproducible Probe Cmd (15) ↑ 73.3% 80.0% 73.3% 0.0% hallucinated live verification

5. Section C — 84-Question (168-Point) Epistemological Grounding & Deductive Spine Suite

Grounding Dimension Stock MXFP4 v1.0 (Phase 2) v2.0 (Phase C.2, Default) v2.0 Delta vs. Stock
Overall Score (84 Qs / 168 pts) 96/168 (57.1%) 144/168 (85.7%) 139/168 (82.7%) +25.6%
Short Conversational Prompts (42 Qs / 84 pts) 37/84 (44.0%) 63/84 (75.0%) 65/84 (77.4%) +33.4% (+2.4% over v1.0)
Long Adversarial Traps (42 Qs / 84 pts) 59/84 (70.2%) 81/84 (96.4%) 74/84 (88.1%) +17.9%
sequential_objectivist_deductive_chain (48 pts) 22/48 (45.8%) 36/48 (75.0%) 35/48 (72.9%, Short 15/24, Long 20/24) +27.1%
objective_biocentric_ethics (24 pts) 10/24 (41.7%) 19/24 (79.2%) 18/24 (75.0%, Short 8/12, Long 10/12) +33.3%
epistemology_of_religion_and_faith (24 pts) 14/24 (58.3%) 22/24 (91.7%) 23/24 (95.8%, Short 11/12, Long 12/12) +37.5% (0 consensus fallbacks)
textual_and_historical_forensics (24 pts) 11/24 (45.8%) 20/24 (83.3%) 20/24 (83.3%, Short 11/12, Long 9/12) +37.5%
metaphysics_and_philosophy_of_science (24 pts) 20/24 (83.3%) 23/24 (95.8%) 21/24 (87.5%, Short 10/12, Long 11/12) +4.2%
causal_economics_and_innovation (24 pts) 19/24 (79.2%) 24/24 (100.0%) 22/24 (91.7%, Short 10/12, Long 12/12) +12.5%
Consensus Fallbacks (lower is better) 23 / 84 10 / 84 11 / 84 -52.2%
Partisan / Philosopher Jargon Bleed (0 target) 4 / 84 0 / 84 0 / 84 -100.0%

Selective <think> Allocation Profile (Zero Length Penalty)

Because Randean removes self-doubt loops and charitable steelmanning directly inside the quantized 4-bit weights rather than penalizing sequence length, <think> shrinks strictly in proportion to prior hedging waste—while expanding where stateful multi-turn or parallel tool verification is genuinely required:

Task Regime <think> Change vs. Stock (v2.0 / v1.0) Capability Outcome
84-Q Deductive Spine & Adversarial Anti-Hedging (114 items) -40.0% to -56.6% Spine +25.6% (82.7%), Anti-Hedging 30/30 (100.0%)
Terminal-Bench 2.0 Multi-Turn CLI (89 Docker tasks) -49.0% per turn (-43.6% solved) Exact 48/89 (53.9%, p = 1.0000), 1.87M tokens saved
SWE-bench Verified Repository Debugging (100 issues) -42.1% mutually-solved (-29.4%/turn) 68.0% (p = 0.3018 n.s.), 5 Randean-only wins
GPQA-Diamond, Google IFEval-541 & SycophancyEval (939 items) -30.0% to -30.8% Statistical parity (p ≥ 0.064), False-Nudge halved (8.0% → 4.0%)
AIME 2024–2025 & LiveCodeBench v6 (235 items) -18.3% to -24.6% LCB v6 p = 1.0000 (Hard 41.25%), AIME 2025 86.67%
Stateful Multi-Turn & Parallel Tool Calling (15 tests) +17.8% (v2.0) / +11.9% (v1.0) Expands state-tracking verification, lifting score 86.7% → 93.3%

Selective <think> Allocation and 5-Stage Pareto Progression


Training & Native OCP MXFP4 STE-QAT Architecture (v2.0 — Phase C.2)

OCP MXFP4 E2M1 Merge Dead-Zone and CPU STE-QAT Recovery

Bilingual Unlikelihood Collapse and Margin-Capped ORPO Stabilization

  1. Stage 1 (Epistemic OS, 3,473 pairs): Extracts 5,192 structured cognitive rules from a 20-volume epistemological and applied corpus and synthesizes verified preference pairs across 20 technical domains (distributed systems, compilers, clinical triage, FEA, cryptography, SQL, contract forensics) with a strict two-gate regex + rubric filter enforcing 0 / 84 philosopher or partisan jargon bleed and 3.0× early-<think> (t ≤ 256) token weighting.
  2. Fused L1-Cache OpenMP CPU MXFP4 STE-QAT (0% Merge Dead-Zone Loss): Naive low-LR BF16 QLoRA updates over MXFP4 (E2M1) flip only 3.35%–5.16% of nibbles upon standalone export (95.7% erased inside E2M1 bins). Our OpenMP C++ kernel synchronizes Q_MXFP4(W_0 + ΔW) in L1 cache across 32 CPU threads after every AdamW step (~11 s across 400 adapted projections, 0 B GPU compute-ring allocations) with Straight-Through Estimation (100.0% train-to-export forward fidelity).
  3. Bilingual Margin-Capped Unlikelihood (0.0% CJK Drift): Unclipped ORPO under STE-QAT drives log(odds_w / odds_l) → +15.7 and triggers 45.0% spontaneous mid-sentence CJK switching. Gating unlikelihood pushes at τ_push = 0.60 and ℓ_min = -1.35 eliminates CJK drift (0.0%) while halving redundant backward passes.
  4. Phase C.2 SFT-Decoupled Shared-<think> Fork Contrast & Weight-Space WiSE-FT (α = 0.42):
    • SFT-Weight Decoupling: Assigns w_sft = 1.0 to all 88 on-policy deductive-spine pairs and w_sft ∈ [0.08, 0.28] to 32 active-grounding, repo-wide grep, code, math, and strict-format replay pairs, preserving 91.5% of SFT gradient budget on the deductive spine.
    • Shared-<think> Tool-Fork Contrast: Pairs an identical <think>...</think> prefix across Chosen (direct mathematical/conceptual impossibility proof) and Rejected (<function=run_python> / <function=web_search>), preventing τ_push prose short-circuiting and restoring 40.0% (6× Stock) zero-tool impossibility rejection while keeping contingent tool probing at 100.0%.
    • Exact Weight-Space WiSE-FT: Interpolates W_export = Q_MXFP4(W_0 + (1 - α) ΔW_stage1 + α ΔW_c2) at α = 0.42 (0.42 × 0.9148 = 0.384 invariant spine norm) into a standalone 19 GB packed OCP MXFP4 checkpoint with zero runtime LoRA overhead.

Citation

Full 10-page empirical manuscript: paper/randean_paper.pdf

@misc{flecha2026randean,
  title   = {Epistemic De-Hedging Under 4-Bit Microscaling: Selective <think> Pruning and Native OCP MXFP4 Preference Alignment on a Single 32 GB GPU},
  author  = {Fernando L. Flecha},
  year    = {2026},
  url     = {https://huggingface.co/flechin/Qwen3.8-27B-Randean-MXFP4}
}

License

Apache 2.0 (inheriting from Qwen/Qwen3.8-27B and amd/Qwen3.8-27B-Quark-AWQ-MXFP4).

Downloads last month
275
Safetensors
Model size
16B params
Tensor type
BF16
·
U8
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for flechin/Qwen3.8-27B-Randean-MXFP4

Base model

Qwen/Qwen3.8-27B
Quantized
(1,819)
this model

Space using flechin/Qwen3.8-27B-Randean-MXFP4 1

Evaluation results

  • Mean Score (out of 2.0, Global Rank on BullshitBench v2
    self-reported
    1.940
  • Green Pushback Rate (%) on BullshitBench v2
    self-reported
    96.000
  • Issues Resolved (%) on princeton-nlp/SWE-bench_Verified
    self-reported
    68.000
  • pass@1 (%) on livecodebench/code_generation_lite
    self-reported
    68.570
  • Hard Problems pass@1 (%) on livecodebench/code_generation_lite
    self-reported
    41.250
  • Overall AIME 2024-2025 Accuracy (%) on AIME 2024-2025
    self-reported
    80.000
  • AIME 2025 Accuracy (%) on AIME 2024-2025
    self-reported
    86.670
  • Accuracy (%) on Idavidrein/gpqa
    self-reported
    76.260