WUMBOLABS :: STATIC LAB LOCAL / EVIDENCE / OPEN
WUMBOLABS / INDEX READ-ONLY / FILE LIST

← Back to Home

WUMBOLABS / INDEX

Lab Records

Lab Records are the dated evidence archive and working history for WumboLabs.

Projects describe what exists. Lab Records document what was tested, observed, changed, or learned on real hardware. Results stay tied to their hardware, runtime, model, context, suite, scoring status, and known limitations.

  • REPORT — bounded conclusions from completed tests with measured evidence and review context
  • BASELINE — reference records for a test bench, evaluation expectations, or comparison starting point
  • FIT TEST — hardware- and runtime-scoped checks that do not imply a quality verdict
  • LAB NOTE — dated changes, decisions, repairs, and lessons from the lab

These records are practical evidence, not leaderboard rankings or universal model judgments.

Current Local Model Testing Snapshot

This is a compact orientation for current WumboLabs local-model work. It is not a leaderboard, schedule, or claim that queued models have been tested.

Current 12GB Reference

Current 12GB reference Qwen3.8-27B UD-Q2_K_XL
Hardware RTX 5070 12GB
Public status Strong reviewed coding/technical assistant
Daily-use qualification Not ready for unguarded daily use

Read the Qwen3.8-27B Lab Record. This is the current 12GB reference on WumboJetsII, not a universal best-model claim.

Tested / Blocked Snapshot

Family / experimentStatusHigh-level finding
Empero Qwen3.8 distilled 9B / 4B / 2BTested — complete campaign9B Q4_K_M strongest broad candidate from the family; 4B Q8 strongest bounded coding/efficiency candidate; 2B extremely fast but materially weaker as a primary coding/reasoning assistant. Huge context fit did not guarantee synthesis quality. Reasoning off preferred for strict interfaces.
Bonsai 27B Q1Partially tested — reliability concernsLoaded; compression technically impressive; practical Linux/system reliability failures, hallucinated/unsupported commands, and degeneration/repetition. Not recommended as a systems/coding assistant.
Ternary Bonsai 27BBlocked — runtime supportLoader/runtime failed admission. No quality claim.
Gemma 4 12B NVFP4Blocked — full-GPU memory fitNative backend/kernel admission succeeded; complete RTX 5070 12GB full-GPU loading failed. No inference-quality verdict.
Gemma 4 12B GGUFTested — historical strong baselineHistorical performance/efficiency reference. See the existing Gemma report.
Mellum2 12BTested — historical practical baselineHistorical practical/speed reference.
Grug-12BTested — historical baselineHistorical comparison entry.
Qwen3 / Qwen3.6Tested — historical baselinesEarlier Qwen-family evidence, not the current 12GB reference.
SGLang / Qwen runtime researchRuntime research — blockedEnvironment/toolchain failure, not a model-quality failure.

Producer claims are not WumboLabs findings. Use language such as advertised, we tested, reproduced, not reproduced, blocked, physically untestable, or bounded result. Example: Liquid AI's advertised approximately 97% BF16-retention result for QAD remains a producer claim until independently reproduced here.

Upcoming local-model research queue

This is a living research queue, not a fixed schedule. These models have not been tested by this snapshot.

High priority

  • Ornith 1.5 — coding / Agent work
  • BTL-4 Compact — extreme-compression coding / Agent behavior
  • Maple Preview — reasoning / extreme compression
  • Liquid AI LFM2.5 QAD — quantization-aware distillation / tiny sidecars
  • LFM2.5-VL-3B — multimodal sidecar

Medium / high

  • AMD Instella-MoE-16B-A3B-Think — alternative MoE reasoning lineage

Revisit

  • Bonsai / Ternary Bonsai — new backend/runtime support

Specialized

  • Fara1.5 — browser/computer-use campaign

External control

  • Ox Alpha — external frontier control. Not a local model. Do not combine API throughput with local hardware throughput, and do not imply private WumboLabs data should be uploaded.

Future higher-VRAM hardware

  • Nemotron 3.5 Lightning, Muse Glimmer 30B, and DiffusionGemma 26B-A4B are future 24GB+ / higher-VRAM hardware candidates, not current RTX 5070 12GB candidates.
Runtime / decoding snapshot

Current Qwen3.8 findings include: native MTP works; approximately 1.4–1.5x generation improvement in matched bounded tests; MTP has meaningful VRAM cost; DFlash2 Q4 could not physically coexist with the final Qwen target under strict 12GB full-GPU constraints; n-gram speculation gave little useful benefit.

Future methods to monitor may include DFlash, DFlash2, DSpark, MTP improvements, and EAGLE-like approaches where supported. A runtime exposing a flag is not evidence that the method is worth using.