Sunnylink Wiki

Extremely new release (Apr 10). Community is currently polling for feedback. No definitive consensus yet, but represents the latest iteration of Off-Policy training. Watch the forum for early feedback. Initial reports on OP Model 8 noted high-speed twitchiness; it remains to be seen if OP 9 resolves this.

Early community feedback is predominantly negative. Users report twitchy, oscillating steering at high speeds, difficulty maintaining lane in stop-and-go traffic, and consistent left-lane hugging. The model appears to struggle with the lateral stability that OP Model 7 had improved. Left-hugging is a consistent issue. If trying this model, adjust camera offset (+0.05 to +0.10). Not recommended for daily driving at this stage.

Significant leap forward for the Off-Policy series. Praised as the most confident OP model yet, with smoother longitudinal deceleration for stopped leads and improved lateral control that self-corrects even when overshooting sharp 90-degree turns. However, random left-hugging persists for some users. Best lateral of the series so far. If hugging left, adjust camera offset (+0.05).

Very early model drop with limited community feedback so far. Initial votes lean positive (Good) but sample size is extremely small. Needs more testing and community evaluation before drawing conclusions. Very few votes so far — check back for updated consensus as more users test it.

Highly polarizing release with growing negative sentiment. Longitudinal control is praised by some (smooth stops behind leads on GV60), but multiple reports of 'afraid of green lights' behavior and failure to stay stopped at red lights. Lateral unwind is improved but stiffer steering causes oscillations and motion sickness on some vehicles. Very vehicle-dependent — F-150 Lightning reports twitchy low-speed behavior. Vehicle dependent. Great on GV60, twitchy on F-150 Lightning. Subaru Crosstrek users report motion sickness. Torque Tune v0 recommended.

Community sentiment has warmed significantly with more testing. Lateral control is vehicle-dependent — Honda Civic users report the best lat in a while with no ping-pong and centering 'like it owes the lane lines money,' while Hyundai Kona and others experience extreme line-to-line ping-ponging above 50mph. More willing to reach set speeds than Pop v2. Still known for creeping at red lights. Very vehicle-dependent. Honda/Kia users trending positive. Torque tune version matters — test with v0 first.

Fifth iteration of off-policy training with significantly more community feedback now. Improved cornering compared to v4 — takes turns wider and away from curbs. Good performance at higher speeds (above 30mph). However, majority rate it as just 'OK' with notable regressions: low-speed behavior struggles to commit to lanes, hugs or crosses lines, and turns lock out partway through. Significant ping-ponging at low speeds reported on Prius Prime. High-speed wobbles above 65mph persist. Community consensus has shifted — majority finds it mediocre. Cornering improvement over v4 but low-speed and turn handling are worse. Torque tune v0 recommended.

Fourth iteration of off-policy training with improved oversteer handling compared to v2. Excellent lane centering and smooth city driving when paired with SCC-V. However, shows hesitation when following a lead vehicle and tends to drive slower than surrounding traffic when no lead is present. Mixed results on curves — improved for Hyundai/Toyota but Tesla users report inside-line riding. SCC-V strongly recommended for safe curve speed handling. Hesitant acceleration when following lead vehicles is the main drawback vs WMI/CD210.

Highly praised third iteration of off-policy training. Community describes it as a 'mad max' version of CD210/WMI — fast longitudinal response with better lateral control than its predecessors. Includes PlanPlus integration for customizable lane centering aggressiveness. Rated A+ by early testers. Minor inconsistency in lead gap following and occasional wobbles at highway speed. PlanPlus integration allows tuning of lane centering vs curve hugging.

Second iteration of off-policy training. Lateral performance is praised — doesn't hug the left side and less huggy on highway turns at 70+. However, longitudinal control is a notable regression: late braking, rubberbanding behind lead cars, and exaggerated follow distances are commonly reported. Best on comma 4; mixed results on other platforms. Strong lateral improvements over v1, but long control needs work. Community opinion is split — some love the lat, others can't tolerate the long behavior.

Experimental model using off-policy training — learns from corrections rather than just cloning. Better highway centering and quicker lateral recentering vs WMI V12. However, experimental long mode feels jerky, and sharp turns can be a step backward. Good potential for highway-focused drivers. 'Off-policy' means it learned from data where the model's behavior differed from the human driver's action, allowing it to learn from corrections more effectively than standard cloning.

Based on WMI V12 architecture with subtle differences. Shows improved red light stopping behavior and less ping-pong at low speeds. However, multiple users report pronounced oversteering on turns — especially problematic on Toyota vehicles with low-torque EPS. Hard braking at close range and red light creeping are also noted drawbacks. Vehicle-dependent: praised on Hyundai Ioniq 5, but struggles with Toyota RAV4 Prime. Consider enabling 'pause lateral while turning' if oversteering is an issue.

The current flagship and daily driver for most users in 2026. Highest intelligence for complex scenarios like intersections, stop signs, and tight urban turns. Can feel slightly 'loose' on straight highways.

Slightly more stable than V12 on straightaways. Good fallback if V12 feels too jittery. Preferred for mostly-highway commutes.

Well-tested mid-generation model. A solid choice if V11/V12 have issues on your specific vehicle.

Likely 'Space Cat' or 'Stable Control' variant. Testing new stability algorithms. Worth trying if you want something different.

Rock-solid choice for newer 2025+ platforms (e.g., Kia K4). Currently the default pre-installed model on Comma 4 hardware. Favored for its stability and predictive behavior on modern ADAS architectures.

Early 2026 build that established the foundation. Still functional for conservative E2E users.

New Year's release that kick-started the 2026 WMI series. Historical significance.

World Model tuned for STIFFER steering response. If WMI V12 feels too 'wiggly', this locks it down. Great for vehicles with loose steering racks.

Late 2025 iteration bridging to 2026 series. More stable than early WMI versions.

Mid-development World Model. Well-tested by the community.

Earlier stable World Model. Less intelligent but proven reliable.

Third iteration of World Model. Foundation for stable releases.

The 'unforgiving' model, refined. Tighter following distance, faster reactions, sharper turns. Excellent for larger vehicles that need authority. Named 'Dark Souls' because it's unforgiving.

Testing a new lateral planner algorithm. Very experimental - for developers and testers.

Second World Model Iteration. Early E2E experiment that showed promise.

The ORIGINAL World Model Iteration. Started the E2E revolution in Sunnypilot.

Original 'unforgiving' model. Still popular for assertive driving. Less refined than v2 but raw character.

Lateral planner experiment. Predecessor to Planplus v2.

Model presented at NeurIPS AI conference. Showcase technology with academic focus.

Original conference model. Interesting for research purposes.

Short-term planning model. Focused on immediate path decisions.

Also referred to as gWM10. A highly discussed and used model noted for its lateral performance. However, some users report it 'hugs the left too much' compared to its predecessor. A popular alternative for those looking for different steering characteristics on GWM platforms.

Generic World Model variant created by the community.

Predecessor to Cool GWM. Well-regarded for its lateral control consistency on supported platforms.

Generic World Model version 9. Standard performance.

Highly praised community model known for its exceptional smoothness and lack of low-speed jerkiness. Many users describe it as 'awesome' and superior to standard releases for lateral control consistency. Particularly popular with Kia e-Niro users in the UK. One of the best custom lateral models of late 2025.

Third iteration of the community-driven model. Active development.

Second version of the community model. Improved on feedback.

Meme-named experimental training run. Inconsistent but fun. Don't use for daily driving.

Updated version of the stable benchmark. Refined lane centering.

Early stable benchmark. Established the 'neutral' baseline.

Original community-driven model. 'By the people, for the people.'

Generic Neural Model version 8. Solid baseline performance.

Generic Neural Model version 6. Less refined than v8.

Generic Neural Model version 5. Mid-stage development.

The 'neutral' benchmark. If a new model feels weird, Nevada is what you compare it to. Well-balanced between smooth and responsive.

Named after aircraft control systems. Smooth, filtered feel like flying. Minimal steering wheel movement.

The raw, unfiltered developer feed from comma.ai's State-of-the-Art (SOTA) training pipeline. This is the absolute 'bleeding edge' of openpilot intelligence, updating almost daily as it ingests new data from the global firehose. HIGH RISK, HIGH REWARD: It often contains the latest breakthroughs in path planning and vision, but has not yet undergone the long-term validation required for a standard release. Technical Note: This model is a direct output of continuous training. Expect radical shifts in behavior day-to-day. If you experience a 'regressive' build, simply wait for the next daily drop or revert to a stable series.

Early Generic Neural Model. Historical interest.

Experimental longitudinal (gas/brake) behavior. Unique acceleration and braking feel.

Custom tune that removes the 'robotic' feeling. Very natural, human-like steering inputs. Community favorite for natural driving feel.

Experimental merge. Unique steering some users love.

Merge of 'Unlimited Vision' and 'Down to Ride.' Combines visibility with comfort.

The legendary 'Wife Approved' model. Sacrifices perfect lane centering for passenger comfort. Won't jerk the wheel; gently drifts into turns. Note: Some Comma 4 users (specifically Ioniq 5) report pronounced right-side lane hugging. Perfect for road trips. If hugging right on C4, try Adjust Camera Offset (-0.05 to -0.10).

Space-themed smooth model with gentle transitions and a relaxed feel.

Earlier Space Lab iteration. Good baseline comfort.

CULT CLASSIC! Often called 'Trash Folder 16.' BEST curve handling of any 2025 model. Drives 'on rails' through winding backroads.

French-named variant. An experimental fork with unique characteristics.

Predecessor to TR16. Good curve handling but slightly less refined.

Previous DTR version. Still smooth and comfortable, slightly less refined than v6.

Simplified planning model. Reliable and predictable behavior.

Standard TR14 iteration. Laid groundwork for TR15 and TR16.

Aggressive tune of the TR series. Combines curve handling with assertive inputs.

Earlier TR series. Foundation for curve-handling characteristics.

Gentle braking profile optimized for hybrids. Maximizes regenerative braking.

Experimental TR-era model with unique steering characteristics.

Named after the butter brand - 'Smooth as butter.' Very gentle inputs.

The name says it all - a hidden gem! Surprisingly good all-around performance.

Early TR series. Foundation for later versions.

Sunnypilot-specific tune. Optimized for this fork.

Original 'fish' model. Smooth braking profile. Historical significance.

Essential for bypassing the factory 20 MPH SCC minimum on modern Hyundais, enabling smooth stop-and-go in heavy traffic.

Customized for Ford's BlueCruise hardware logic. Excellent torque management for heavy trucks and SUVs.

The Torque Control Profile V3. Highly regarded for older Honda platforms to prevent steering micro-vibrations.

Maintained by MVL. Specifically targets the steering deadzone issues present in the Honda Odyssey platform.

Optimized for comma 3X hardware on TSS2 Toyotas. Drastically improves behavior in dense city grids.

Solid legacy model. Named after the cathedral - built to last.

Known for SUPER STIFF steering. Hugs center lane and REFUSES to move. On rails. Can feel robotic but incredibly stable.

Named because it 'lubricated' North Dakota's stiffness. Best of both worlds. Stable centering without robotic feel.

Legacy comfort model with very soft inputs. Passenger-focused.

THE most reliable legacy model. The 'Toyota Camry' of models - boring, predictable, rock solid. UNSHAKEABLE on highways. Bad at stop signs.

Predecessor to Recertified Herbalist. Still good but less polished.

The 'VISION KING' of older models. EXCELLENT in rain or poor lane markings. Can see lanes that other models miss.

Very old model from 2023. Mostly deprecated. Historical interest only. Don't use for daily driving.